Multi-cable vibration identification method and system based on physical information generation adversarial network

Abstract

The application provides a multi-cable vibration identification method and system based on a physical information generative adversarial network, and relates to the technical field of cable identification.The method provided by the application comprises the following steps: constructing a conditional generative adversarial network, taking random noise and a condition vector as inputs, and outputting a two-dimensional simulation vibration field; using a physical constraint loss function composed of a wave equation residual loss term, a frequency domain physical consistency loss term, a time domain physical consistency loss term and a cross-cable coupling consistency loss term to jointly optimize and train the generator with an adversarial loss; converting the simulation vibration field into a time-frequency feature sequence, inputting the time-frequency feature sequence into a neural network model to extract a cable sample feature vector; constructing a standard feature vector database of each type of cable, and determining the cable identity by calculating the cosine similarity between the feature vector of a to-be-measured signal and the standard feature vectors in the database.The application effectively improves the problem that the specific cable identity cannot be identified in a multi-cable co-site scene due to the scarcity of real fault samples and the distortion of simulation data.

Description

Technical Field

[0001] This invention relates to the field of optical cable identification technology, and in particular to a method and system for identifying vibrations in multiple optical cables based on physical information generative adversarial networks. Background Technology

[0002] In industrial sectors such as power, telecommunications, and rail transportation, phase-sensitive optical time-domain reflectometry (Φ-OTDR) distributed acoustic sensing (DAS) technology is a crucial method for monitoring infrastructure safety. This technology analyzes Rayleigh scattering signals in optical fibers to detect and locate vibrations and contact events along optical cables in real time. In real-world scenarios, such as urban underground utility tunnels or power transmission towers, multiple optical cables are often laid closely together. When these areas are subjected to external disturbances such as construction, excavation, or vehicle traffic, existing systems can only determine the range of vibration but cannot pinpoint the specific cable from which it originates. In other words, the system can only report vibration at a location without specifying which cable is affected, forcing maintenance personnel to conduct on-site inspections of each cable, resulting in slow response times, increased maintenance costs, and potential safety hazards.

[0003] The difficulties in identification stem from two main factors. Firstly, in environments with multiple optical cables coexisting, the weak characteristic signals of the target cable are easily masked by strong environmental noise, power interference, and signals from other optical cables. Traditional signal processing methods struggle to reliably extract distinguishable feature information from this mixed signal, hindering high-precision identification. Secondly, existing methods typically use simulations based on simplified physical formulas to generate optical cable vibration data. However, such data is overly idealized and fails to reflect the complexity and randomness of noise in real-world environments, as well as the true coupling relationship between signals and interference. This results in unstable performance and limited identification capabilities for processing methods based on such data in practical applications. More critically, real fault samples are virtually impossible to obtain in actual engineering projects, and the data quality generated by existing simulation methods is insufficient to support subsequent processing. Existing vibration signal generation methods based on generative adversarial networks (GANs) disclose schemes for synthesizing vibration signals through conditional GANs and comparing the time-domain and frequency-domain features of the generated signals. However, their physical parameters are only used as conditional inputs and posterior evaluation indicators, without imposing differential operator constraints at the level of partial differential equations describing the propagation of optical cable vibrations on the generator output, nor addressing the coupling consistency problem of generated signals among multiple co-located optical cables. Another type of method based on distributed optical fiber vibration intelligent monitoring discloses vibration signal acquisition and identification processes in multiple scenarios, but it does not address the physical consistency synthesis of vibration data, nor does it solve the problem of identifying multiple co-located optical cables under conditions of scarce real fault samples. In summary, the technical challenge faced by existing technologies is that the scarcity of real fault samples and the distortion of simulation data prevent the realization of specific optical cable identification in multi-optical cable co-location scenarios. Therefore, it is urgent to develop a solution to address the above problems. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for identifying vibrations in multiple optical cables based on physical information generative adversarial networks, which improves the problem in the prior art that the scarcity of real fault samples and the distortion of simulation data make it impossible to identify the specific optical cable in a multi-optical cable co-location scenario.

[0005] The multi-fiber cable vibration identification method and system based on physical information generative adversarial networks provided by this invention adopts the following technical solution:

[0006] Firstly, a multi-fiber cable vibration identification method based on a physical information generative adversarial network specifically includes:

[0007] A conditional generative adversarial network is constructed. The generator of the conditional generative adversarial network takes a random noise vector and a conditional vector as input and outputs a two-dimensional simulated vibration field along the spatial position dimension and time dimension of the optical cable. The conditional vector includes the fundamental frequency, damping coefficient, undamped natural angular frequency, propagation speed of mechanical disturbance along the optical cable, and physical spacing between the co-located optical cables.

[0008] A physical constraint loss function is constructed using the residual loss term of the wave equation, the frequency domain physical consistency loss term, the time domain physical consistency loss term, and the cross-optical cable coupling consistency loss term. This physical constraint loss function is then jointly optimized with the adversarial loss to train the generator, generating a two-dimensional simulated vibration field that simultaneously satisfies the intrinsic physical parameter constraints of the target optical cable and the coupling law between co-located optical cables. Specifically, the residual loss term of the wave equation, through automatic differentiation, constrains the two-dimensional simulated vibration field to satisfy the one-dimensional damped wave equation at the differential operator level. The frequency domain physical consistency loss term constrains the peak frequency of the generated signal to match the fundamental frequency or harmonic structure. The time domain physical consistency loss term constrains the envelope attenuation characteristics of the generated signal to match the damping coefficient. The cross-optical cable coupling consistency loss term constrains the normalized cross-correlation between the generated signals of co-located optical cables to conform to the exponential coupling attenuation law determined by the optical cable spacing.

[0009] The two-dimensional simulated vibration field is divided into single-optical-cable simulated vibration signals along the spatial position dimension, converted into time-frequency feature sequences, and then input into a neural network model. The neural network model includes a convolutional layer, a bidirectional long short-term memory layer, and an attention layer. The attention layer performs weighted aggregation on the features of each time step output by the bidirectional long short-term memory layer to output an optical cable sample feature vector.

[0010] A standard feature vector database is constructed using the arithmetic mean of the feature vectors of at least two samples of optical cables for each type of optical cable.

[0011] The cosine similarity between the optical cable sample feature vector of the vibration signal to be identified and each standard feature vector in the standard feature vector database is calculated, and the optical cable identity is determined based on the cosine similarity.

[0012] This invention provides a multi-optical cable vibration identification method based on a physical information generative adversarial network (GAN). By constructing a conditional GAN, the generator takes random noise and a conditional vector containing the fundamental frequency, damping coefficient, undamped natural angular frequency, propagation velocity, and co-location distance of at least two co-located optical cables as input. The generator outputs a two-dimensional simulated vibration field unfolded along the spatial and temporal dimensions of the optical cables. The generator output is constrained by a physical constraint loss function composed of wave equation residual loss terms, frequency domain physical consistency loss terms, time domain physical consistency loss terms, and cross-optical cable coupling consistency loss terms. This function, combined with the adversarial loss function, is used to jointly optimize and generate a method that simultaneously satisfies the intrinsic physical parameters of the target optical cable. A two-dimensional simulated vibration field is used to determine the coupling law between bundled and co-located optical cables. The simulated vibration field is segmented along the spatial position and converted into a time-frequency feature sequence. The sequence is then input into a neural network model containing convolutional layers, long short-term memory layers, and attention layers to extract feature vectors. The arithmetic mean of the feature vectors of at least two similar samples is used as the standard feature vector for each type of optical cable to construct a standard feature vector database. During identification, the cosine similarity between the feature vector of the signal under test and each standard feature vector in the standard feature vector database is calculated to determine the identity of the optical cable. This improves the problem in the existing technology that the scarcity of real fault samples and the distortion of simulation data make it impossible to identify the specific optical cable in a multi-optical cable co-location scenario.

[0013] Optionally, the frequency domain physical consistency loss term constrains the estimated peak frequency of the generated signal to match the fundamental frequency or harmonic structure by using a local window spectrum centroid method centered on the fundamental frequency corresponding to the fundamental frequency in the condition vector.

[0014] Optionally, the temporal physical consistency loss term constrains the envelope decay characteristics of the generated signal to match the damping coefficient by using the logarithmic domain error between the moving root mean square envelope of the generated signal and the theoretical exponential decay curve determined by the damping coefficient in the conditional vector.

[0015] Optionally, the calculation process of the residual loss term of the wave equation is as follows: along the time dimension and spatial position dimension of the two-dimensional simulated vibration field, the time second-order partial derivative, the time first-order partial derivative, and the spatial second-order partial derivative of the two-dimensional simulated vibration field are calculated by finite difference or inverse automatic differentiation; the results of multiplying the time second-order partial derivative and the time first-order partial derivative with the damping term coefficient determined by the damping coefficient and the undamped natural angular frequency in the condition vector, and the results of multiplying the spatial second-order partial derivative with the square of the propagation velocity, are added in the form of a one-dimensional damped wave equation to obtain the equation residual of the two-dimensional simulated vibration field at each spatiotemporal sampling point; the mean of the square of the norm of the equation residual at all spatiotemporal sampling points is used as the loss value of the residual loss term of the wave equation.

[0016] Optionally, the calculation process of the cross-optical cable coupling consistency loss term is as follows: the two-dimensional simulated vibration field is divided into two or more single-optical cable simulated vibration signals along the spatial position dimension; the mean value of any two single-optical cable simulated vibration signals is removed and normalized along the time dimension, and the normalized cross-correlation function value is calculated; the target coupling attenuation curve value is constructed based on the physical distance between co-located optical cables and the coupling attenuation coefficient determined by the co-located medium in the condition vector; the mean square error between the normalized cross-correlation function value of all co-located optical cables and the target coupling attenuation curve value is used as the loss value of the cross-optical cable coupling consistency loss term.

[0017] Optionally, the physical constraint loss function further includes a physical parameter cycle consistency loss term; the physical parameter cycle consistency loss term is implemented by introducing a physical parameter estimation network parallel to the discriminator. The physical parameter estimation network takes the two-dimensional simulated vibration field output by the generator or the simulated vibration signal of a single optical cable obtained by slicing along the spatial position dimension as input, and outputs the estimated values ​​of the fundamental frequency and damping coefficient corresponding to the generated signal; the mean square error between the estimated value and the true value of the corresponding fundamental frequency and damping coefficient in the condition vector is used as the loss value of the physical parameter cycle consistency loss term, constraining the generator to retain the physical parameter semantics specified by the condition vector while deceiving the discriminator.

[0018] Optionally, converting the simulated vibration signal of the single optical cable into a time-frequency feature sequence includes: performing a short-time Fourier transform on the simulated vibration signal of the single optical cable to obtain a complex time spectrum, extracting the amplitude spectrum of the complex time spectrum, scaling the processed amplitude spectrum to a preset fixed size using an interpolation algorithm, and stacking the interpolated amplitude spectra in chronological order to form the time-frequency feature sequence.

[0019] Optionally, when the attention layer performs weighted aggregation on the features of each time step output by the bidirectional long short-term memory layer to output the optical cable sample feature vector, it includes: calculating the energy score of the hidden state output by the bidirectional long short-term memory layer through a fully connected layer and a hyperbolic tangent activation function, then normalizing it through a Softmax function to obtain the attention weight, and using the attention weight to weightedly sum the hidden states of each time step to obtain the optical cable sample feature vector.

[0020] Optionally, the construction of the standard feature vector database includes: using a trained neural network model to extract optical cable sample feature vectors from at least two samples of each type of optical cable, grouping them according to optical cable category, calculating the arithmetic mean of each group, and using it as the standard feature vector for that type of optical cable.

[0021] Secondly, a multi-fiber optic cable vibration identification system based on a physical information generative adversarial network specifically includes:

[0022] The data synthesis module is used to construct a conditional generative adversarial network. The generator of the conditional generative adversarial network takes a random noise vector and a conditional vector as input and outputs a two-dimensional simulated vibration field along the spatial position dimension and time dimension of the optical cable. The conditional vector includes the fundamental frequency, damping coefficient, undamped natural angular frequency, propagation speed of mechanical disturbance along the optical cable, and physical spacing between the co-located optical cables.

[0023] The loss constraint module is used to construct a physical constraint loss function using the residual loss term of the wave equation, the frequency domain physical consistency loss term, the time domain physical consistency loss term, and the cross-optical cable coupling consistency loss term. This physical constraint loss function is then jointly optimized with the adversarial loss to train the generator, generating a two-dimensional simulated vibration field that simultaneously satisfies the intrinsic physical parameter constraints of the target optical cable and the coupling law between co-located optical cables. Specifically, the residual loss term of the wave equation, through automatic differentiation, constrains the two-dimensional simulated vibration field to satisfy the one-dimensional damped wave equation at the differential operator level; the frequency domain physical consistency loss term constrains the peak frequency of the generated signal to match the fundamental frequency or harmonic structure; the time domain physical consistency loss term constrains the envelope attenuation characteristics of the generated signal to match the damping coefficient; and the cross-optical cable coupling consistency loss term constrains the normalized cross-correlation between the generated signals of co-located optical cables to conform to the exponential coupling attenuation law determined by the optical cable spacing.

[0024] The model building module is used to divide the two-dimensional simulated vibration field into single-optical cable simulated vibration signals along the spatial position dimension, convert them into time-frequency feature sequences, and then input them into a neural network model. The neural network model includes a convolutional layer, a bidirectional long short-term memory layer, and an attention layer. The attention layer performs weighted aggregation on the features of each time step output by the bidirectional long short-term memory layer to output an optical cable sample feature vector.

[0025] The data storage module is used to construct a standard feature vector database by using the arithmetic mean of the feature vectors of at least two optical cable samples of each type of optical cable as the standard feature vector.

[0026] The signal recognition module is used to calculate the cosine similarity between the optical cable sample feature vector of the vibration signal to be identified and each standard feature vector in the standard feature vector database, and to determine the identity of the optical cable based on the cosine similarity.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] First, this invention transforms the physical generation mechanism of optical cable vibration signals into training constraints that can participate in the backpropagation of neural networks. A physical constraint loss function is constructed using differentiable wave equation residual loss terms, frequency domain physical consistency loss terms, time domain physical consistency loss terms, and cross-optical cable coupling consistency loss terms. This function is directly embedded into the generator's training process, rather than performing conventional time-frequency analysis or post-processing screening on the generated signals. Specifically, the wave equation residual loss term performs second-order automatic differentiation on the two-dimensional simulated vibration field output by the generator, substituting the resulting partial derivatives into the one-dimensional damped wave equation to construct the equation residual. This constrains the generated signals to satisfy the partial differential equations of optical cable vibration propagation at the differential operator level. The cross-optical cable coupling consistency loss term ensures the co-located physical consistency of the generated dataset at the multi-optical cable coupling level by constraining the normalized cross-correlation between signals generated from co-located optical cables to conform to the exponential coupling attenuation law determined by the optical cable spacing, thus ensuring the co-located physical consistency of the generated dataset at the multi-optical cable coupling level. Compared with ordinary GANs, this invention can effectively suppress spectral position drift, temporal envelope distortion, and non-physical pseudo-samples that do not satisfy the propagation law of wave equation; compared with pure physical simulation, it can also use the discriminator to learn environmental noise and complex coupling features from real field data, effectively solving the problem of identification difficulties caused by the scarcity of real fault samples and the distortion of simulation data.

[0029] Second, this invention achieves separation of feature extraction and classification decision through a decoupled recognition architecture of a standard feature vector database and cosine similarity matching. Unlike end-to-end classifiers that require retraining when adding new categories, this invention only needs to calculate the standard feature vector of the new category and add it to the database when adding a new optical cable category, without retraining the feature extraction network, thus exhibiting good scalability and deployment flexibility.

[0030] Third, the present invention also uses an attention layer to weighted aggregate the features of each time step output by the long short-term memory layer, so that the model can automatically identify and suppress time steps dominated by environmental common-mode interference, and focus on time steps containing physical fingerprint features of optical cables, thereby maintaining stable recognition performance in complex environments with strong noise and multiple co-located optical cables. Attached Figure Description

[0031] Figure 1A flowchart of a multi-optical cable vibration identification method based on physical information generative adversarial networks provided in an embodiment of the present invention;

[0032] Figure 2 A structural diagram of a multi-fiber optic cable vibration identification system based on a physical information generative adversarial network provided in an embodiment of the present invention;

[0033] Figure 3 A comparison of time-frequency diagrams of different optical cable physical characteristics in embodiments of the present invention;

[0034] Figure 4 This is a comparison diagram of the response of the same optical cable under different interference events in an embodiment of the present invention;

[0035] Figure 5 This is a comparison diagram of the responses of different optical cables to the same excavation interference event in an embodiment of the present invention;

[0036] Figure 6 This is a t-SNE dimensionality reduction distribution diagram of the optical cable feature vector in an embodiment of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.

[0038] See Figure 1 This invention provides a method for identifying vibrations in multiple optical cables based on a physical information generative adversarial network, comprising the following steps:

[0039] S1. Construct a conditional generative adversarial network. The generator of the conditional generative adversarial network takes a random noise vector and a conditional vector as input and outputs a two-dimensional simulated vibration field along the spatial position dimension and time dimension of the optical cable. The conditional vector includes the fundamental frequency, damping coefficient, undamped natural angular frequency, propagation speed of mechanical disturbance along the optical cable, and physical spacing between the co-located optical cables.

[0040] S2. A physical constraint loss function is constructed using the residual loss term of the wave equation, the frequency domain physical consistency loss term, the time domain physical consistency loss term, and the cross-optical cable coupling consistency loss term. This physical constraint loss function is then jointly optimized with the adversarial loss to train the generator, generating a two-dimensional simulated vibration field that simultaneously satisfies the intrinsic physical parameter constraints of the target optical cable and the coupling law between co-located optical cables. Specifically, the residual loss term of the wave equation, through automatic differentiation, constrains the two-dimensional simulated vibration field to satisfy the one-dimensional damped wave equation at the differential operator level. The frequency domain physical consistency loss term constrains the peak frequency of the generated signal to match the fundamental frequency or harmonic structure. The time domain physical consistency loss term constrains the envelope attenuation characteristics of the generated signal to match the damping coefficient. The cross-optical cable coupling consistency loss term constrains the normalized cross-correlation between the generated signals of co-located optical cables to conform to the exponential coupling attenuation law determined by the optical cable spacing.

[0041] S3. The two-dimensional simulated vibration field is divided into single optical cable simulated vibration signals along the spatial position dimension, converted into time-frequency feature sequences, and then input into a neural network model. The neural network model includes a convolutional layer, a bidirectional long short-term memory layer, and an attention layer. The attention layer performs weighted aggregation on the features of each time step output by the bidirectional long short-term memory layer to output an optical cable sample feature vector.

[0042] S4. Construct a standard feature vector database by taking the arithmetic mean of the feature vectors of at least two samples of optical cables for each type of optical cable as the standard feature vector;

[0043] S5. Calculate the cosine similarity between the optical cable sample feature vector of the vibration signal to be identified and the standard feature vector in the standard feature vector database, and determine the identity of the optical cable based on the cosine similarity.

[0044] In some embodiments, when constructing the conditional generative adversarial network in step S1, the input to its generator consists of two parts: a random noise vector that follows a standard normal distribution. and the conditional vector of the intrinsic physical parameters of the encoded optical cable In this invention, the intrinsic physical parameters of an optical cable refer to a set of physical properties that are determined by the cable's material, structure, laying condition, and environmental constraints, and can be used to distinguish different individual optical cables. These properties include, but are not limited to, fundamental frequency, damping coefficient, harmonic structure, dispersion coefficient, cutoff frequency, and group refractive index. These parameters guide the generator to produce simulated vibration signals of a single optical cable with specific physical fingerprints.

[0045] In one specific embodiment, the condition vector Including the baseband of the target optical cable and damping coefficient It can also further include intrinsic physical parameters such as harmonic structure and dispersion coefficient. Conditional Generative Adversarial Networks The generator uses a random noise vector that follows a standard normal distribution. and conditional vectors characterizing different physical properties of optical cables As input, the output is a two-dimensional simulated vibration field unfolded along the spatial and temporal dimensions of the optical cable. .

[0046] The fundamental difference between this embodiment and existing GAN-based optical cable vibration signal generation schemes lies in the following: Existing schemes typically only learn the statistical distribution of the real signal at the generation result level, or perform conventional time-frequency analysis on the generated signal before applying posterior screening, without embedding the partial differential equations describing the optical cable vibration mechanism as differentiable constraints into the generator training loop. This embodiment directly calculates the second-order partial derivatives of the two-dimensional simulated vibration field output by the generator along both the spatial and temporal dimensions, substitutes the obtained partial derivatives into the one-dimensional damped wave equation to construct the equation residuals, and uses the mean square of the equation residuals as the loss term to participate in the generator parameter update. This method differs from any form of time-frequency analysis, envelope fitting, or analytical solution curve comparison; it applies physical mechanism constraints to the generated signal at the differential operator level.

[0047] This embodiment expands the output dimension of the generator, making the generator output the spatial position dimension along the optical cable. and time dimension Two-dimensional simulated vibration field Instead of a one-dimensional time-series signal at a single point, this structural change makes it computationally feasible to impose partial differential equation residual constraints on the generator output: in each training iteration, using the finite difference or automatic differentiation mechanism of the deep learning framework, the generator output is constrained by partial differential equations. Calculate the second-order partial derivative of time along the time index and along the spatial index. Second-order partial derivatives in space Substituting the aforementioned second-order partial derivatives into the one-dimensional damped wave equation determined by the target optical cable damping coefficient, undamped natural angular frequency, and propagation velocity, the equation residuals are constructed. The mean squared norm of the equation residuals at all spatiotemporal sampling points is used as the residual loss term of the wave equation. Together with the adversarial loss, frequency domain physical consistency loss term, and time domain physical consistency loss term, the generator parameters are optimized.

[0048] In some embodiments, in step S2, the physical constraint loss function is constructed using the residual loss term of the wave equation, the frequency domain physical consistency loss term, the time domain physical consistency loss term, and the cross-optical cable coupling consistency loss term. This physical constraint loss function is then jointly optimized with the adversarial loss to train the generator, generating a two-dimensional simulated vibration field that simultaneously satisfies the intrinsic physical parameter constraints of the target optical cable and the coupling law between co-located optical cables. Essentially, this physical constraint loss function is a concrete mathematical representation of the optical cable wave equation and the Rayleigh scattering statistical model. It should be noted that the purpose of constructing the frequency domain physical consistency loss term and the time domain physical consistency loss term in this embodiment is not to perform a general time-frequency analysis of the input signal, but rather to transform the physical laws of optical cable vibration into differentiable constraints during the generator training process. If only parameters such as the fundamental frequency and damping coefficient are used as post-processing judgment indicators, the generator may still produce pseudo-samples during training that are similar to the real data distribution but do not conform to the optical cable vibration propagation law, such as main frequency shift, energy peaks induced by noise, or envelope attenuation that is too fast or too slow. To avoid the aforementioned problems, this embodiment maps the target fundamental frequency and damping coefficient to frequency-domain physical consistency loss and time-domain physical consistency loss, respectively, and these losses, along with the adversarial loss, are applied to the generator's parameter updates, ensuring that the generator is constrained by the physical fingerprint of the target optical cable in each iteration. Specifically, the free decaying vibration of the optical cable after external excitation follows a one-dimensional damped wave equation in macroscopic mechanics. Separating its dominant fundamental frequency mode can be equivalent to a second-order linear ordinary differential equation, the specific form of which is as follows:

[0049] ,

[0050] in, This represents the optical cable's vibration displacement as a function of time. Represents the undamped natural angular frequency, satisfying ; Let represent the damping coefficient of the target optical cable. The theoretical analytical solution to this equation is:

[0051] ,

[0052] in, This represents the initial amplitude determined by the intensity of the external excitation. This represents the damped natural angular frequency, which satisfies the relationship with the undamped natural angular frequency. ; Represents a time variable; The initial phase is determined by the initial physical state.

[0053] Based on the theoretical analytical solution, this invention naturally decouples it into two parts: frequency domain resonance characteristics and time domain damping attenuation law. On the one hand, the cosine oscillation term in the theoretical solution... This determines that the signal has an extreme energy peak in the frequency domain. Specifically, a frequency domain physical consistency loss term is constructed. The peak frequency used to constrain the generated signal spectrum is related to the fundamental frequency in the condition vector. Consistent. Specifically, regarding the generated signal. The amplitude spectrum is obtained by performing a Fast Fourier Transform. Since the traditional maximum index operation is non-differentiable, it can lead to gradient breakage during network training. This embodiment uses a differentiable local window spectral centroid method to extract the peak frequency. The implementation process is as follows: First, the target frequency index is locked based on the target fundamental frequency and frequency resolution, and a local frequency window is extracted centered on this index. Then, the amplitude values ​​of each frequency point within the window are used as weights to perform a weighted summation of the frequency indexes, calculating a continuous and differentiable estimated centroid frequency point. Finally, the estimated peak frequency is calculated. With the target base frequency The difference is used to construct the frequency domain physical consistency loss term. The formula used is as follows:

[0054] ,

[0055] in, This represents a local window range centered on the ideal frequency point corresponding to the target fundamental frequency (e.g., the window radius is preferably 5 frequency points). Indicates the frequency index within a local window; This indicates the frequency resolution of the Fast Fourier Transform; Indicates the amplitude spectrum in the index The amplitude value at that point; Represents a minimal constant to prevent the denominator from being zero (e.g.) The technical advantage of the aforementioned local centroid method lies in its ability to not only eliminate non-differentiable nodes in network backpropagation but also effectively avoid interference from global high-frequency noise, thereby improving the noise resistance of extreme frequency estimation. Furthermore, the local window spectral centroid method used in this embodiment is not for ordinary spectral analysis of the generated signal, but rather for constructing a frequency domain physical consistency loss that can participate in the generator's backpropagation training. Specifically, by restricting the peak frequency estimation to a local frequency window corresponding to the target fundamental frequency and replacing the non-differentiable maximum value indexing operation with a continuously differentiable spectral centroid, the main energy frequency band of the generated signal continuously converges towards the fundamental frequency or harmonic structure of the target optical cable during training. Thus, the generator not only learns the statistical distribution of real vibration data but also maintains physical fingerprint characteristics consistent with the intrinsic physical parameters of the target optical cable in the frequency domain, avoiding the generation of non-physical single-cable simulated vibration signals with main frequency drift or dominated by noise pseudo-peaks.

[0056] On the other hand, the exponential decay term in the theoretical solution This determines the attenuation trajectory of the signal envelope. Simultaneously, a temporal-domain physical consistency loss term is constructed. The envelope attenuation trajectory of the generated signal is constrained to conform to an exponential damping law. Specifically, the envelope of the generated signal is extracted using a sliding root mean square (RMS) window. The window length is preferably configured to 500 sampling points, and the sliding step size is preferably configured to 100 sampling points to ensure that the envelope curve can smoothly filter high-frequency carriers while accurately preserving the low-frequency attenuation trend. After extracting the envelope, the mean square error between the envelope in the logarithmic domain and the theoretical exponential attenuation curve is calculated. The formula used is as follows:

[0057] ,

[0058] in, This indicates that the root mean square envelope of the generated signal is extracted using specified window parameters. This represents the simulated vibration signal of a single optical cable obtained by slicing the two-dimensional simulated vibration field output by the generator along the spatial position dimension. This indicates the window length of the sliding root mean square window; Indicates the sliding step size; This represents the amplitude of the generated signal envelope at the initial moment, used for normalization of the envelope; This represents a small constant to prevent the denominator from being zero and to prevent overflow in logarithmic operations; for example, it can be taken as... ; Indicates the damping coefficient of the target optical cable; Indicates the undamped natural angular frequency of the target optical cable; This represents the time vector corresponding to each envelope sampling point; This indicates the number of envelope sampling points involved in the calculation; This represents the square of the L2 norm. In the above formula, the root mean square envelope of the generated signal is first divided by its initial amplitude to eliminate the influence of differences in the initial amplitude of different simulated vibration signals of single optical cables on the loss calculation; then, the logarithm of the normalized envelope is taken to transform the theoretical exponential decay relationship into an approximately linear relationship. Since the theoretical normalized envelope satisfies... Therefore, by minimizing The squared L2 norm of the signal can constrain the envelope attenuation trajectory of the generated signal to conform to the exponential damping law determined by the target damping coefficient. Among these, This indicates that the sliding root mean square envelope of the generated signal is extracted using the window parameters at time [time value missing]. The value of ; This represents the moving root mean square envelope of the generated signal at the initial time. The corresponding amplitude is used to normalize the envelope to eliminate the influence of initial amplitude differences in different single-cable simulated vibration signals on loss calculation; the temporal physical consistency loss term preferably uses the logarithmic mean square error between the sliding root mean square envelope and the theoretical exponential decay curve. This method does not only compare the signal amplitude, but also extracts the low-frequency decay trend of the generated signal through the sliding root mean square envelope, and constrains this decay trend in the logarithmic domain to be consistent with the theoretical exponential decay law corresponding to the damping coefficient. Since the actual optical cable vibration signal is superimposed with random noise and high-frequency texture, directly constructing the loss on the original time-domain waveform is easily affected by phase deviation and transient noise; this embodiment uses logarithmic domain constraints at the envelope level, which can more stably reflect the damping decay law, so that the generated sample maintains a temporal physical fingerprint related to the target optical cable laying state and material properties.

[0059] Ultimately, the physical constraint loss function is: ,in, and These are the weighting coefficients for the frequency-domain physical consistency loss term and the time-domain physical consistency loss term, respectively, used to balance the optimization intensity of different physical constraints in gradient descent based on their magnitude differences. In some preferred embodiments, and The value range can be set to More preferably, a dynamic normalization allocation strategy based on the current loss value can be adopted (i.e., ... , ,in, This represents a very small constant to prevent the denominator from being zero. This indicates that gradient backpropagation is blocked to adaptively ensure that the magnitudes of the two losses remain consistent throughout the training process. During training, the adversarial loss and the physical constraint loss are jointly optimized. Furthermore, to prevent the frequency-domain physical consistency loss term and the temporal-domain physical consistency loss term from dominating in the long term due to differences in their dimensions and numerical ranges, this embodiment employs a dynamic normalization weight allocation strategy. This strategy adaptively adjusts the corresponding weights based on the current numerical magnitudes of the frequency-domain and temporal losses, and can perform gradient blocking operations on the weight calculation branch, ensuring that the weights are only used to balance the contributions of different loss terms to the generator update without introducing additional backpropagation paths. Through this setting, the frequency-domain constraints and temporal constraints can remain relatively balanced during training, preventing the generator from satisfying only one physical law while ignoring another.

[0060] In addition to the frequency-domain physical consistency loss term and the time-domain physical consistency loss term, this embodiment also constructs a wave equation residual loss term to be output from the constraint generator at the differential operator level. Specifically, the wave equation residual loss term... The calculation formula is:

[0061] ,

[0062] in, The spatial index represents the two-dimensional simulated vibration field output by the generator. Time Index The value at; Indicates the damping coefficient of the target optical cable; Indicates the undamped natural angular frequency of the target optical cable; This indicates the propagation speed of mechanical disturbances along the optical cable, which is determined by the intrinsic physical parameters of the optical cable. , Represents the number of spatial and temporal sampling points, respectively; second-order partial derivatives and It is obtained by finite difference or inverse automatic differentiation.

[0063] The technical advantages of the residual loss term in the wave equation are as follows: First, the constraint object is elevated from "the statistical characteristics of the signal" to "the physical equation satisfied by the signal," so that the generator is subject to hard constraints at the differential operator level of the optical cable vibration propagation law at each parameter update, avoiding non-physical pseudo-samples caused by relying solely on statistical learning; Second, due to the propagation term in the spatial dimension of the one-dimensional damped wave equation... It is defined only in a two-dimensional spatiotemporal field. Therefore, the two-dimensional vibration field output structure of this embodiment has a one-to-one correspondence with the loss term, which is different from any existing scheme that constructs a loss function on a single-point time-series signal. Third, the calculation path of the loss term is entirely composed of finite difference or automatic differentiation, and it shares the same calculation graph with the adversarial loss. It can be backpropagated end-to-end, which is different from existing schemes that use physical parameters as post-processing judgment indicators or a priori filtering conditions.

[0064] Furthermore, addressing the technical problem of difficulty in identifying specific optical cables in multi-cable co-location scenarios, this embodiment introduces a cross-cable coupling consistency loss term during generator training. In typical co-location scenarios such as urban underground utility tunnels and shared tower installations, adjacent optical cables not only exhibit mechanical coupling transmitted through laying media (such as conduits, soil, and cable trays), but also weak coherent crosstalk generated by the superposition of Rayleigh scattering light fields during Φ-OTDR demodulation. Physically, this coupling effect manifests as follows: the cross-correlation of vibration signals between co-located optical cables decreases exponentially with increasing spatial spacing, exhibiting significant coupling in the near field (small spacing) and tending towards decorrelation in the far field (large spacing). Existing GAN-based vibration signal generation schemes train independently for each individual optical cable, failing to constrain the coupling patterns between co-located optical cables during the generation stage, resulting in a lack of physical consistency in the generated multi-cable dataset under co-location scenarios. The generator in this embodiment uses the intrinsic physical parameters and spacing information of multiple co-located optical cables as conditional vectors to output two-dimensional simulated vibration fields of multiple co-located optical cables in parallel. The cross-optical cable coupling consistency loss term The calculation formula is:

[0065] ,

[0066] in, This represents the set of all co-located optical fiber pairs participating in the training. , The two-dimensional simulated vibration field represents the vibration field at the t-th... , No. Simulated vibration signal of a single optical cable obtained from a slice of a spatial location; Indicates the first , The physical spacing between optical cables; This represents the coupling attenuation coefficient determined by the medium in the co-located region; The normalized cross-correlation function of the two signals along the time dimension is expressed by the following formula:

[0067] ,

[0068] in, , These represent the mean values ​​of the corresponding signals along the time dimension. This represents a very small constant to prevent the denominator from being zero. , These represent the corresponding optical cables in time. The vibration amplitude at that location.

[0069] The technical effect of this loss term is twofold: physically, it ensures that the generated multi-fiber optic cable dataset conforms to the propagation law of "near-field coupling and far-field decorrelation" in co-located scenarios; at the recognition level, it allows the downstream feature extraction network to be exposed to a sample distribution that conforms to the co-located coupling law during the training phase, thereby significantly improving the model's ability to distinguish the identity of target fiber optic cables in real multi-fiber optic cable co-located scenarios. Mathematically, this loss term must be constructed based on the joint output of signals generated by multiple fiber optic cables. Schemes that generate signals independently for a single fiber optic cable cannot structurally achieve this constraint, thus distinguishing it from all existing GAN-type vibration signal generation methods that are trained independently for a single fiber optic cable.

[0070] Furthermore, to further avoid the generator losing the semantics of the physical parameters specified by the condition vector simply because it only satisfies the above physical constraints (e.g., the fundamental frequency or damping coefficient may be close to the target, but drifts after being dominated by the adversarial loss), this embodiment can further introduce a physical parameter cycle consistency loss term: a physical parameter estimation network is set up in parallel outside the discriminator. It takes the generator output as input and outputs estimates of the fundamental frequency and damping coefficient, and uses the mean square error between the estimates and the true value of the condition vector as the cycle consistency loss term. This is optimized together with the aforementioned loss terms. This mechanism enables the generative adversarial network to form a closed-loop constraint of reverse "parameter reduction" in addition to the forward "conditional control," preventing the generator from sacrificing the semantics of physical parameters to minimize adversarial loss. The physical parameter cycle consistency loss term... The calculation formula is:

[0071] ,

[0072] in, Indicates that the generator uses noise and condition vector The simulated vibration field generated as input; This represents a physical parameter estimation network configured in parallel with the discriminator. Represents the condition vector The fundamental frequency and damping coefficient subvectors in the data; Indicates the noise and condition vector Find the expected value of the joint distribution.

[0073] In summary, the final physical constraint loss function in this embodiment... Represented as:

[0074] ,

[0075] in, , , , and These represent the weight coefficients of the corresponding loss terms, and a dynamic normalization allocation strategy based on the current loss value can be used to adaptively balance the magnitudes. The final training objective of the generator is... ,in, This represents the adversarial loss of the generator.

[0076] In this embodiment, a two-dimensional simulated vibration field with 16 channels and 4096 time points was trained for 1200 steps using three different random seeds (seed=42, seed=7, seed=2023) to verify the convergence effect of each physical consistency loss term. The results are shown in Table 1 below:

[0077] Table 1 Comparison of convergence effects for each loss term.

[0078] Wave equation residual loss term Approximately 1.19–1.51 Approximately 0.026–0.058 Approximately 46 times / 51 times / 23 times Cross-optical cable coupling consistency loss term Approximately 0.30–0.53 Approximately 0.012 to 0.022 Approximately 45 times / 26 times / 14 times Temporal physical consistency loss term Approximately 0.21 Approximately 0.027 Approximately 7.9 times Frequency domain physical consistency loss term Approximately 0.25 Approximately 0.23 Maintain stability

[0079] Among them, the decrease factors of the wave equation residual loss term and the cross-optical cable coupling consistency loss term under three different initial conditions all exceeded one order of magnitude, indicating that the convergence effect of the newly added physical constraints has stable reproducibility across initial conditions. The above experimental results prove that the generator output converges to the physical consistency determined by the target physical parameters at both the partial differential equation operator level and the co-located multi-optical cable coupling law level.

[0080] The above experimental results verify that during the generator training process, the residual loss term of the wave equation and the cross-optical cable coupling consistency loss term both decreased by more than an order of magnitude, proving that the generator output converges to the physical consistency determined by the target physical parameters at both the partial differential equation operator level and the co-located multi-optical cable coupling law level.

[0081] Specifically, the adversarial loss adopts a standard minimax game form. For the generator, the adversarial loss takes the following form:

[0082] ,

[0083] in, Represents the sampled random noise vector and condition vector Calculate the mathematical expectation; Indicates that the generator uses noise The resonant characteristics also conform to the time-domain damping law and the conditional vector. The simulated vibration field generated as input is divided along the spatial position dimension, and the resulting single optical cable simulated vibration signal is input into the discriminator. This represents the probability value output by the discriminator, used to characterize the discriminator's confidence level in considering the input signal as real field data. The calculation logic and technical effect of this adversarial loss formula are as follows: the generator's optimization objective is to "deceive" the discriminator as much as possible, making it so that the discriminator misclassifies the simulated vibration signal from a single optical cable as a real signal. Approaching 1 (i.e., 100% certain it is true); through the outer wrapping negative logarithmic function When this probability approaches 1, the loss value The probability approaches 0 (reaching minimum loss), but as the probability approaches 0, the loss value approaches positive infinity, resulting in a huge gradient penalty. This design allows the generator to effectively update network parameters using the standard gradient descent algorithm, continuously improving the realism of the generated signal in terms of environmental noise and overall distribution.

[0084] Furthermore, considering that the signals acquired by distributed acoustic sensing not only include macroscopic mechanical vibrations but also the interference effect of coherent Rayleigh backscattering within the optical fiber, the random spatial distribution of the scattering centers in the absence of strong external strain causes the complex amplitude of the backscattered light field to follow a two-dimensional Gaussian distribution, exhibiting strong random statistical characteristics. To simulate the aforementioned Rayleigh scattering statistical model, this invention constructs a physical information generative adversarial network topology. Its generator uses a high-dimensional standard normal distribution random noise vector as one of the input variables. The network learns to couple this random noise with a defined physical condition vector, synthesizing broadband environmental noise and high-frequency harmonic textures while simulating the dominant frequency waveform. In some specific embodiments, to achieve the generation of the aforementioned long-time-series high-frequency vibration signal (e.g., generating a signal with a duration of 5 seconds and a sampling rate of 10000Hz, i.e., a single sample length of 50000 data points), the generator input receives the concatenated features (e.g., concatenating a noise vector of dimension 100 with a condition vector of dimension 2, resulting in a total of 102 input channels), which are then sequentially upsampled through five one-dimensional transposed convolutional layers. Specifically, the number of output channels in layers one through five is mapped sequentially from 128 to 64, 32, 16, and 1, respectively. The first four layers are cascaded with a one-dimensional batch normalization layer and a Leaky ReLU activation function, while the fifth layer cascades a Tanh activation function with the output size as... The time-domain vibration signal of a single optical fiber cable is simulated. Correspondingly, the discriminator uses a multi-layer one-dimensional convolutional network to progressively downsample and extract temporal features for binary classification of true and false signals. Its downsampling feature extraction block contains 4 to 5 one-dimensional convolutional layers symmetrical to the generator structure. Each layer is followed only by a cascaded LeakyReLU activation function without batch normalization to preserve the absolute scale difference of the data. Finally, after flattening, the data is fed into a fully connected layer to output a probability scalar value. The real signal data used by the discriminator for discrimination comes from real environmental and event vibration data collected on the optical fiber cable by a distributed acoustic sensing system deployed in the field.

[0085] Through alternating training using the above mechanisms, the generator attempts to deceive the discriminator while minimizing physical constraint loss, and the discriminator only optimizes adversarial loss to distinguish between real and generated signals. Ultimately, this drives the generator to output a simulated vibration field that satisfies both frequency domain resonance characteristics and time domain damping laws, while the discriminator forces the generated signal to approximate the real data distribution, thereby obtaining high-fidelity training samples that combine physical consistency and environmental realism.

[0086] Using the above method, a total of 15,000 samples were generated for 10 different types of optical cables. Each sample lasted 5.0 seconds and the sampling rate was 10,000 Hz. Figure 3 This is a time-frequency diagram comparison of the physical characteristics of different optical cables (taking the first optical cable and the sixth optical cable as examples) in an embodiment of the present invention. Figure 3 The left side shows the time-frequency diagram of the first optical cable. Figure 3 The right side shows the time-frequency diagram of the sixth optical cable; the horizontal axis of both diagrams is time (seconds), and the vertical axis is frequency (Hz). The color bar on the right indicates the logarithmic power (dB) corresponding to the grayscale values ​​in the time-frequency diagram. The left diagram shows the fundamental frequency energy distribution and harmonic texture structure of the first optical cable generated by the physical information generative adversarial network, showing that its energy concentration band is biased towards the lower frequency band; the right diagram shows the time-frequency characteristics of the sixth optical cable, with a significantly different energy concentration band in its high-frequency range (such as around 500Hz). Figure 3 The left-right comparison shows that the time-frequency diagrams of different optical cables exhibit significantly different fundamental frequency energy distributions and harmonic texture structures, verifying that the data generated by this invention can effectively preserve the unique physical fingerprint characteristics of optical cables.

[0087] In some embodiments, when converting the simulated vibration signal of a single optical cable into a time-frequency feature sequence in step S3, a Short-Time Fourier Transform (STFT) is first performed on the simulated vibration signal of the single optical cable. The window length of the STFT is set to 256 sampling points, the overlap length is 128 sampling points, and the window function can be a Hanning window. After the STFT, a complex time spectrum is obtained, the dimension of which is [number of frequencies, number of time frames].

[0088] Subsequently, the amplitude spectrum of the complex time-frequency spectrum, i.e., the complex modulus value at each time-frequency point, is extracted. A logarithmic transform is then performed on the amplitude spectrum to enhance the contrast of weak signal features. Specifically, this is achieved using... Logarithmic transformation is then performed. Next, Z-score normalization is applied: for each sample's amplitude spectrum, the mean is subtracted and then divided by the standard deviation, resulting in an amplitude spectrum with zero mean and unit variance, thus eliminating differences in amplitude dimensions between different samples. After normalization, bilinear interpolation is used to scale the amplitude spectrum of each time frame to a preset fixed size. For example, scaling to... Pixels, thus meeting the input dimension requirements of subsequent neural network models.

[0089] Then, the simulated vibration signal from a single optical cable is divided into multiple continuous and equal-length time segments, each segment corresponding to a time step. For example, if the total signal duration is 5 seconds, and each segment is 1 second, a total of 5 time steps are obtained. The STFT, amplitude spectrum extraction, logarithmic transformation, normalization, and scaling operations are independently performed on each time step to obtain the two-dimensional amplitude spectrum for that time step. Finally, the amplitude spectra of at least two consecutive time steps are stacked in chronological order to form a three-dimensional time-frequency feature sequence with dimensions [number of time steps, height, width]. For example, when using 5 time steps and each amplitude spectrum being 64×64, the feature sequence dimension is (5, 64, 64).

[0090] In some embodiments, when constructing the neural network model in step S3, the model adopts a cascaded structure, which includes a convolutional layer, a bidirectional long short-term memory layer, and an attention layer in sequence.

[0091] Convolutional layers (CNNs) are used to extract spatial texture features from single-frame time-frequency feature maps. The convolutional layers in this model consist of three cascaded convolutional blocks. In one specific embodiment, each convolutional block sequentially includes: a two-dimensional convolutional layer (e.g., a 3×3 kernel, stride of 1, and padding of 1), a batch normalization layer, a ReLU activation layer, and a max-pooling layer (e.g., a 2×2 window). The first convolutional block can have 32 output channels, the second 64, and the third 128. Through these convolutional blocks, the input time-frequency feature sequence is encoded into a high-dimensional spatial feature sequence.

[0092] Bidirectional Long Short-Term Memory (Bi-LSTM) layers are used to model the forward and backward dependencies of the aforementioned spatial feature sequences in the temporal dimension, specifically capturing the signal attenuation patterns determined by the material damping properties. Let the feature sequence output by the convolutional layer be... ,in The time step is defined. A bidirectional Long Short-Term Memory (LSTM) layer processes the sequence through two LSTM chains, one forward and one backward, outputting the hidden state sequence for each time step. It contains contextual information about the past and future, represented as ,in, This indicates a positive LSTM output. This indicates the inverse LSTM output.

[0093] The attention layer is used to weight and aggregate the hidden states of the bidirectional LSTM output at each time step to suppress time steps affected by environmental common-mode interference and focus on time steps containing the physical fingerprint features of the optical cable. The calculation process of the attention layer is as follows: First, the hidden states of each time step are weighted and aggregated. Input a fully connected layer, pass it through a hyperbolic tangent activation function to obtain the energy fraction. : ,in, This represents the transpose of the learnable context vector. This represents the learnable weight matrix. This represents the bias term. Then, the energy score is input into the Softmax function to obtain the normalized attention weights. ,in, Indicates the first Energy fraction at each time step This represents the total number of time steps. Finally, the hidden states at each time step are weighted and summed using attention weights to obtain a fixed-dimensional feature vector. This feature vector is used as the feature vector of the optical cable sample for subsequent identification.

[0094] Through the above mechanism, the model can automatically extract distinctive optical cable physical features from time-frequency feature sequences and exhibit robustness to environmental interference.

[0095] Based on the time-frequency feature sequence generated in step S3 and the corresponding optical cable identification tags, the above neural network model is trained using supervised learning. In some preferred embodiments, the training configuration is as follows: the adaptive moment estimation optimizer (Adam) is used for parameter updates, and the initial learning rate is set to... The loss function used is the cross-entropy loss function, which calculates the loss value using only the classification confidence score (Logits) obtained from the model output layer and the true label; the training is conducted for 30 epochs; after each epoch, the accuracy is evaluated on the validation set, and the model parameters with the highest accuracy on the validation set are saved as the final best model.

[0096] In some embodiments, when constructing the standard feature vector database in step S4, the neural network model trained and optimized in step S3 is first loaded, and all learnable parameters of the model are fixed.

[0097] Then, all simulation samples are iterated. For each simulation sample, it is converted into a time-frequency feature sequence using the same time-frequency feature sequence construction method as in step S3, input into the neural network model, and the feature vector of the sample is obtained from the output of the attention layer. Record the feature vector of each sample and its corresponding optical cable identification label.

[0098] Group all simulation samples according to their fiber optic cable identification tags. Assume there are a total of... Different types of optical cables, the first Optical fiber cables include Each of the n samples corresponds to a feature vector. For the nth sample... For fiber optic cables of a certain type, the arithmetic mean of the feature vectors of all its samples is calculated, and this is used as the standard feature vector for that type of fiber optic cable. The calculation formula is as follows:

[0099] ,

[0100] in, Indicates the first Standard feature vectors for optical fiber cables; For the first Total number of samples of optical fiber cables; Indicates the first A sample index set for optical fiber cables; Indicates the first The feature vectors extracted from each sample using a neural network model.

[0101] Finally, all standard feature vectors and their corresponding fiber optic cable identification tags are associated and stored to form a query database. This database can be stored in the memory of an electronic device or in an external storage medium for subsequent inference.

[0102] In some embodiments, in step S5, the cosine similarity between the optical cable sample feature vector of the vibration signal to be identified and each standard feature vector in the standard feature vector database is calculated. When determining the optical cable identity based on the cosine similarity, firstly, the vibration signal to be identified undergoes the same preprocessing procedure as in step S3 to obtain the feature vector. Then, the standard feature vectors for each type of optical cable are read from the database constructed in step S4. The cosine similarity between the feature vector of the signal to be identified and each standard feature vector is calculated. All cosine similarity values ​​are compared, and the category index corresponding to the maximum value is selected as the final identification result. The formula used is as follows:

[0103] .

[0104] Figure 4This is a comparison of the response of the same optical cable (taking the fourth optical cable as an example) under different interference events in an embodiment of the present invention. The left, middle, and right columns correspond to the excavation event, the drilling event, and the vehicle event, respectively. The top of each column is a time-frequency plot, and the bottom is the normalized attention weight distribution. Taking the excavation event in the left column as an example: the time-frequency plot shows that the interference signal manifests as a short-duration broadband impact with extremely high energy at approximately 2.0 seconds; observing the corresponding attention distribution below, it can be seen that the model automatically reduces the attention weight to an extremely low level at the instant of the strong impact (at 2.0 seconds), while allocating the highest weight to the relatively stable period before the interference outbreak (approximately 0.5 to 1.5 seconds), and retaining the secondary weight for a short period after the impact ends (approximately 2.5 seconds). For the drilling event and the vehicle event, although their time-frequency manifestations are continuous broadband noise of different forms, the model can still dynamically adjust the attention, concentrating the high weight on the period that is less affected by external noise and can reflect the intrinsic characteristics of the optical cable. Figure 4 The model demonstrates that it can automatically identify and isolate high-energy regions dominated by environmental common-mode interference, adaptively focus on optical cable segments with high signal-to-noise ratio, and possess decoupling and anti-interference capabilities.

[0105] Figure 5 This is a comparison of the responses of different optical cables (first, fourth, and sixth optical cables) to the same mining interference event in an embodiment of the present invention. The top row shows the time-frequency plots of each optical cable, and the bottom row shows the corresponding normalized attention weight distribution. The time-frequency plots show that although all cables are subjected to the same mining impact at approximately 2.0 seconds, the fundamental frequency energy band and harmonic textures in the low-frequency range of different optical cables still exhibit significant differences due to their different physical parameters. More importantly, observing the attention distribution in the bottom row reveals that although the frequency domain textures of the input signals are different, the attention evolution trajectories allocated by the model to different optical cables are highly consistent. At the moment of the mining impact at 2.0 seconds, the attention weights of all three types of optical cables are strongly suppressed by the model; while the highest attention weights are uniformly anchored in the 0.5 to 1.5 second range, and the secondary weight distribution at 2.5 seconds also remains synchronized. Figure 5 This demonstrates that the attention-based anti-interference mechanism of this invention is not a rigid overfitting to specific optical cable characteristics, but rather learns the general rules of actively suppressing external common-mode mutations and adaptively extracting internal differential-mode physical fingerprints, thus possessing cross-optical cable generalization capabilities.

[0106] Figure 6 This is a t-SNE dimensionality reduction distribution diagram of the optical cable feature vector in an embodiment of the present invention. Figure 6 The image shows the distribution of feature vectors for ten different types of optical cables (optical cables 1 to 10) in two-dimensional space. Points with different geometric shapes represent different optical cable categories. The horizontal axis represents the t-SNE dimensionality reduction component 1, and the vertical axis represents the t-SNE dimensionality reduction component 2. Figure 6As can be seen, the feature points of the ten types of optical cables form ten clusters with clear boundaries, no overlap and high cohesion. The features of the same type of optical cable are closely clustered, while the features of different types are kept at a significant distance. Figure 6 This demonstrates that the physical fingerprints extracted by this invention have extremely high intra-class consistency and inter-class separability.

[0107] refer to Figure 2 This invention provides a multi-fiber optic cable vibration identification system based on a physical information generative adversarial network, comprising the following steps:

[0108] The data synthesis module 100 is used to construct a conditional generative adversarial network. The generator of the conditional generative adversarial network takes a random noise vector and a conditional vector as input and outputs a two-dimensional simulated vibration field along the spatial position dimension and time dimension of the optical cable. The conditional vector includes the fundamental frequency, damping coefficient, undamped natural angular frequency, propagation speed of mechanical disturbance along the optical cable, and physical spacing between the co-located optical cables.

[0109] The loss constraint module 200 is used to construct a physical constraint loss function using the residual loss term of the wave equation, the frequency domain physical consistency loss term, the time domain physical consistency loss term, and the cross-optical cable coupling consistency loss term. This physical constraint loss function is then jointly optimized with the adversarial loss to train the generator, generating a two-dimensional simulated vibration field that simultaneously satisfies the intrinsic physical parameter constraints of the target optical cable and the coupling law between co-located optical cables. Specifically, the residual loss term of the wave equation, through automatic differentiation, constrains the two-dimensional simulated vibration field to satisfy the one-dimensional damped wave equation at the differential operator level. The frequency domain physical consistency loss term constrains the peak frequency of the generated signal to match the fundamental frequency or harmonic structure. The time domain physical consistency loss term constrains the envelope attenuation characteristics of the generated signal to match the damping coefficient. The cross-optical cable coupling consistency loss term constrains the normalized cross-correlation between the generated signals of the co-located optical cables to conform to the exponential coupling attenuation law determined by the optical cable spacing.

[0110] The model building module 300 is used to divide the two-dimensional simulated vibration field into single optical cable simulated vibration signals along the spatial position dimension, convert them into time-frequency feature sequences, and then input them into a neural network model. The neural network model includes a convolutional layer, a bidirectional long short-term memory layer, and an attention layer. The attention layer performs weighted aggregation on the features of each time step output by the bidirectional long short-term memory layer to output an optical cable sample feature vector.

[0111] The data storage module 400 is used to construct a standard feature vector database by using the arithmetic mean of the feature vectors of at least two optical cable samples of each type of optical cable as the standard feature vector.

[0112] The signal recognition module 500 is used to calculate the cosine similarity between the optical cable sample feature vector of the vibration signal to be identified and each standard feature vector in the standard feature vector database, and to determine the identity of the optical cable based on the cosine similarity.

[0113] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.

Claims

1. A method for identifying vibration in multiple optical cables based on physical information generative adversarial networks, characterized in that, include: A conditional generative adversarial network is constructed. The generator of the conditional generative adversarial network takes a random noise vector and a conditional vector as input and outputs a two-dimensional simulated vibration field along the spatial position dimension and time dimension of the optical cable. The conditional vector includes the fundamental frequency, damping coefficient, undamped natural angular frequency, propagation speed of mechanical disturbance along the optical cable, and physical spacing between the co-located optical cables. A physical constraint loss function is constructed using the residual loss term of the wave equation, the frequency domain physical consistency loss term, the time domain physical consistency loss term, and the cross-optical cable coupling consistency loss term. This physical constraint loss function is then jointly optimized with the adversarial loss to train the generator, generating a two-dimensional simulated vibration field that simultaneously satisfies the intrinsic physical parameter constraints of the target optical cable and the coupling law between co-located optical cables. Specifically, the residual loss term of the wave equation, through automatic differentiation, constrains the two-dimensional simulated vibration field to satisfy the one-dimensional damped wave equation at the differential operator level. The frequency domain physical consistency loss term constrains the peak frequency of the generated signal to match the fundamental frequency or harmonic structure. The time domain physical consistency loss term constrains the envelope attenuation characteristics of the generated signal to match the damping coefficient. The cross-optical cable coupling consistency loss term constrains the normalized cross-correlation between the generated signals of co-located optical cables to conform to the exponential coupling attenuation law determined by the optical cable spacing. The calculation process of the cross-optical cable coupling consistency loss term is as follows: the two-dimensional simulated vibration field is divided into two or more single-optical-cable simulated vibration signals along the spatial position dimension. For any two single-optical-cable... The simulated vibration signal of the cable is mean-normalized along the time dimension, and the normalized cross-correlation function value is calculated. A target coupling attenuation curve value is constructed based on the physical spacing between co-located optical cables and the coupling attenuation coefficient determined by the co-located medium in the condition vector. The mean square error between the normalized cross-correlation function value and the target coupling attenuation curve value of each pairwise combination of all co-located optical cables is used as the loss value of the cross-optical cable coupling consistency loss term. The physical constraint loss function also includes a physical parameter cyclic consistency loss term. The physical parameter cyclic consistency loss term is implemented by introducing a physical parameter estimation network parallel to the discriminator. The physical parameter estimation network takes the two-dimensional simulated vibration field output by the generator or the single-cable simulated vibration signal obtained by slicing along the spatial position dimension as input, and outputs estimated values ​​of the fundamental frequency and damping coefficient corresponding to the generated signal. The mean square error between the estimated value and the true values ​​of the corresponding fundamental frequency and damping coefficient in the condition vector is used as the loss value of the physical parameter cyclic consistency loss term, constraining the generator to retain the physical parameter semantics specified by the condition vector while deceiving the discriminator. The two-dimensional simulated vibration field is divided into single-optical-cable simulated vibration signals along the spatial position dimension, converted into time-frequency feature sequences, and then input into a neural network model. The neural network model includes a convolutional layer, a bidirectional long short-term memory layer, and an attention layer. The attention layer performs weighted aggregation on the features of each time step output by the bidirectional long short-term memory layer to output an optical cable sample feature vector. A standard feature vector database is constructed using the arithmetic mean of the feature vectors of at least two optical cable samples of each type as the standard feature vector. The cosine similarity between the optical cable sample feature vector of the vibration signal to be identified and each standard feature vector in the standard feature vector database is calculated, and the optical cable identity is determined based on the cosine similarity.

2. The method as described in claim 1, characterized in that, The frequency domain physical consistency loss term constrains the estimated peak frequency of the generated signal to match the fundamental frequency or harmonic structure by using a local window spectrum centroid method centered on the fundamental frequency corresponding to the fundamental frequency in the condition vector.

3. The method as described in claim 1, characterized in that, The temporal physical consistency loss term constrains the envelope decay characteristics of the generated signal to match the damping coefficient by using the logarithmic domain error between the moving root mean square envelope of the generated signal and the theoretical exponential decay curve determined by the damping coefficient in the conditional vector.

4. The method as described in claim 1, characterized in that, The calculation process of the residual loss term of the wave equation is as follows: along the time dimension and spatial position dimension of the two-dimensional simulated vibration field, the time second-order partial derivative, the time first-order partial derivative, and the spatial second-order partial derivative of the two-dimensional simulated vibration field are calculated by finite difference or inverse automatic differentiation; the results of multiplying the time second-order partial derivative and the time first-order partial derivative with the damping term coefficient determined by the damping coefficient and the undamped natural angular frequency in the condition vector, and the results of multiplying the spatial second-order partial derivative with the square of the propagation velocity are added in the form of a one-dimensional damped wave equation to obtain the equation residual of the two-dimensional simulated vibration field at each spatiotemporal sampling point; the mean of the square of the norm of the equation residual at all spatiotemporal sampling points is used as the loss value of the residual loss term of the wave equation.

5. The method as described in claim 1, characterized in that, When converting the simulated vibration signal of the single optical cable into a time-frequency feature sequence, the process includes: performing a short-time Fourier transform on the simulated vibration signal of the single optical cable to obtain a complex time spectrum, extracting the amplitude spectrum of the complex time spectrum, scaling the processed amplitude spectrum to a preset fixed size using an interpolation algorithm, and stacking the interpolated amplitude spectra in chronological order to form the time-frequency feature sequence.

6. The method as described in claim 1, characterized in that, When the attention layer performs weighted aggregation of the features output by the bidirectional long short-term memory layer at each time step to output the optical cable sample feature vector, it includes: calculating the energy score of the hidden state output by the bidirectional long short-term memory layer through a fully connected layer and a hyperbolic tangent activation function, then normalizing it through a Softmax function to obtain the attention weight, and using the attention weight to weightedly sum the hidden states at each time step to obtain the optical cable sample feature vector.

7. The method as described in claim 1, characterized in that, The construction of the standard feature vector database includes: using a trained neural network model to extract the optical cable sample feature vectors of at least two samples for each type of optical cable, grouping them according to the optical cable category, calculating the arithmetic mean of each group, and using it as the standard feature vector for that type of optical cable.

8. A multi-fiber optic cable vibration identification system based on physical information generative adversarial networks, characterized in that, include: The data synthesis module is used to construct a conditional generative adversarial network. The generator of the conditional generative adversarial network takes a random noise vector and a conditional vector as input and outputs a two-dimensional simulated vibration field along the spatial position dimension and time dimension of the optical cable. The conditional vector includes the fundamental frequency, damping coefficient, undamped natural angular frequency, propagation speed of mechanical disturbance along the optical cable, and physical spacing between the co-located optical cables. The loss constraint module is used to construct a physical constraint loss function using the residual loss term of the wave equation, the frequency domain physical consistency loss term, the time domain physical consistency loss term, and the cross-optical cable coupling consistency loss term. This physical constraint loss function is then jointly optimized with the adversarial loss to train the generator, generating a two-dimensional simulated vibration field that simultaneously satisfies the intrinsic physical parameter constraints of the target optical cable and the coupling law between co-located optical cables. Specifically, the residual loss term of the wave equation, through automatic differentiation, constrains the two-dimensional simulated vibration field to satisfy the one-dimensional damped wave equation at the differential operator level. The frequency domain physical consistency loss term constrains the peak frequency of the generated signal to match the fundamental frequency or harmonic structure. The time domain physical consistency loss term constrains the envelope attenuation characteristics of the generated signal to match the damping coefficient. The cross-optical cable coupling consistency loss term constrains the normalized cross-correlation between the generated signals of co-located optical cables to conform to the exponential coupling attenuation law determined by the optical cable spacing. The calculation process of the cross-optical cable coupling consistency loss term involves dividing the two-dimensional simulated vibration field along the spatial position dimension into two or more single-optical-cable simulated vibration signals. The simulated vibration signals of two single optical cables are mean-normalized along the time dimension, and the normalized cross-correlation function value is calculated. A target coupling attenuation curve value is constructed based on the physical spacing between co-located optical cables and the coupling attenuation coefficient determined by the co-located medium in the condition vector. The mean square error between the normalized cross-correlation function value and the target coupling attenuation curve value of all co-located optical cables is used as the loss value of the cross-optical cable coupling consistency loss term. The physical constraint loss function also includes a physical parameter cyclic consistency loss term. This physical parameter cyclic consistency loss term is implemented by introducing a physical parameter estimation network parallel to the discriminator. The physical parameter estimation network takes the two-dimensional simulated vibration field output by the generator or the simulated vibration signal of a single optical cable obtained by slicing along the spatial position dimension as input, and outputs estimated values ​​of the fundamental frequency and damping coefficient corresponding to the generated signal. The mean square error between the estimated value and the true values ​​of the corresponding fundamental frequency and damping coefficient in the condition vector is used as the loss value of the physical parameter cyclic consistency loss term, constraining the generator to retain the physical parameter semantics specified by the condition vector while deceiving the discriminator. The model building module is used to divide the two-dimensional simulated vibration field into single-optical cable simulated vibration signals along the spatial position dimension, convert them into time-frequency feature sequences, and then input them into a neural network model. The neural network model includes a convolutional layer, a bidirectional long short-term memory layer, and an attention layer. The attention layer performs weighted aggregation on the features of each time step output by the bidirectional long short-term memory layer to output an optical cable sample feature vector. The data storage module is used to construct a standard feature vector database by using the arithmetic mean of the feature vectors of at least two optical cable samples of each type of optical cable as the standard feature vector. The signal recognition module is used to calculate the cosine similarity between the optical cable sample feature vector of the vibration signal to be identified and each standard feature vector in the standard feature vector database, and to determine the identity of the optical cable based on the cosine similarity.

Images