Long distance ofdr fast demodulation method and device based on training neural network with synthetic data

Abstract

This invention discloses a method and apparatus for fast demodulation of long-distance OFDR based on synthetic data-trained neural networks. It achieves rapid prediction of spectral drift by constructing generative training samples that closely match the real characteristics of long-distance OFDR and combining them with a neural network model. The method includes: generating multiple training sample data with long-distance transmission degradation characteristics in batches using a data synthesis method according to the physical characteristics of Rayleigh scattering signals in optical fibers; using the training sample data to supervise the training of a preset neural network model to obtain a trained drift prediction model; acquiring local reference signals and local measurement signals of each sensing region on the sensing fiber using an optical frequency domain reflectometer; constructing corresponding inference sample data based on the local reference signals and local measurement signals of each sensing region, and sequentially inputting the inference sample data of each sensing region into the drift prediction model for forward inference, then receiving the spectral drift prediction values ​​of each sensing region sequentially output by the drift prediction model; and converting the spectral drift prediction values ​​of each sensing region into corresponding predicted values ​​of the measured parameters according to a pre-calibrated inversion function between spectral drift and the measured parameter.

Description

Technical Field

[0001] This invention relates to distributed optical fiber sensing technology, and more particularly to a method and apparatus for fast demodulation of long-distance OFDR based on synthetic data-trained neural networks. Background Technology

[0002] Optical frequency domain reflectometer (OFDR) is one of the core technologies in distributed fiber optic sensing systems. With its ultra-high spatial resolution, high detection sensitivity, and long-distance monitoring potential, it has become the mainstream technique for detecting physical parameters in distributed fiber optics. Changes in the environment in which the fiber is located, as well as changes in its physical state such as deformation and strain, can cause frequency or spectral shifts in the Rayleigh scattering spectrum within the fiber. OFDR technology compares the Rayleigh scattering signals of the fiber under reference and measurement conditions, extracts the spectral shift between the reference and measurement signals, and then combines this with calibration relationships to invert and obtain various measured parameters such as distributed temperature and strain along the fiber.

[0003] In OFDR signal processing, compensating for the nonlinear tuning error of the swept-frequency laser and effectively mapping the frequency domain signal to the spatial domain signal is a prerequisite for accurately extracting the spectral drift. For short-distance fiber optic monitoring scenarios, the industry commonly uses an auxiliary interferometer to perform phase resampling and interpolation correction on the main interferometer signal to obtain a Rayleigh scattering signal with high signal-to-noise ratio and high coherence. This is then combined with traditional matching algorithms such as cross-correlation to solve for the spectral drift, achieving ideal detection results. However, the correction capability of the auxiliary interferometer is limited by the length of its internal delay fiber. Once the actual sensing distance exceeds this limit, the interpolation resampling method based on the auxiliary interferometer will no longer be applicable and cannot meet the needs of long-distance monitoring.

[0004] For ultra-long-distance OFDR monitoring scenarios, the current mainstream solution is to use deskewing filtering algorithms to process the master interferometer signal, achieving signal mapping from the frequency domain to the spatial domain under long-distance conditions. However, as the sensing distance increases, laser phase noise accumulates continuously. Combined with phase noise estimation errors and higher-order term truncation errors present in the deskewing filtering process, this leads to decreased coherence of the Rayleigh scattering signal at the far end of the fiber, increased noise amplitude, and degraded signal morphology. Under these conditions, the effective correlation characteristics between the reference and measurement signals are significantly weakened. Traditional sliding cross-correlation algorithms require interval-by-interval sliding matching and peak retrieval, resulting in cumbersome computational processes and high overhead. Furthermore, in low-coherence, high-noise environments, problems such as correlation peaks being submerged by noise, signal mismatch, and peak finding failure easily occur, ultimately leading to a significant decrease in detection accuracy and operational stability in the far-end region.

[0005] To address the shortcomings of traditional algorithms, data-driven technologies have been gradually introduced into the field of OFDR signal processing. Neural networks can learn the nonlinear correlation between reference and measurement signals through sample training, exhibiting excellent feature extraction capabilities in complex noisy environments. After training, they can output spectral drift results simply through forward propagation, with computational efficiency far exceeding that of traditional cross-correlation algorithms that rely on point-by-point retrieval. This makes them more suitable for the real-time processing of massive amounts of long-distance fiber optic data.

[0006] However, existing OFDR detection schemes based on neural networks still have significant drawbacks: On the one hand, the model performance of such methods is highly dependent on the quality and distribution of training samples. If the training samples do not fully simulate the real characteristics such as coherence degradation, noise interference, amplitude fluctuations, and local signal fading caused by long-distance deskewing filtering, it will cause a mismatch between the distribution of training data and the field measured data, resulting in high noise, insufficient accuracy, and poor stability in long-distance remote areas. On the other hand, the collection, calibration, and manual annotation of long-distance OFDR measured data are difficult and costly, making it difficult to obtain a large number of measured samples covering different drift degrees, different noise intensities, and different signal degradation states, which further limits the application of neural network technology in long-distance OFDR systems. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, this invention provides a method and apparatus for rapid demodulation of long-distance OFDR based on synthetic data-trained neural networks. By constructing generative training samples that fit the real characteristics of long-distance OFDR, and combining them with a neural network model, it achieves rapid prediction of spectral drift, balancing detection accuracy, computational efficiency, and long-distance adaptability. It can effectively complete the distributed detection of various physical parameters such as fiber temperature and strain.

[0008] The technical problem to be solved by the present invention is achieved through the following technical solution: A fast demodulation method for long-range OFDR based on synthetic data-trained neural networks includes the following steps: Step 1: Based on the physical characteristics of fiber Rayleigh scattering signals, multiple training sample data with long-distance transmission degradation characteristics are generated in batches using data synthesis. Each training sample data includes a local reference analog signal, a local measurement analog signal and its corresponding spectral drift label. Step 2: Use the training sample data to perform supervised training on the preset neural network model to obtain a drift prediction model that has completed training; Step 3: Acquire the local reference real signal and local measurement real signal of each sensing area on the sensing fiber using an optical frequency domain reflectometer; Step 4: Based on the local reference real signal and local measurement real signal of each sensing area, construct corresponding inference sample data, and input the inference sample data of each sensing area into the drift prediction model for forward inference in sequence. Then, receive the spectral drift prediction value of each sensing area output by the drift prediction model in sequence. Step 5: Based on the inversion function between the pre-calibrated spectral drift and the measured parameter, convert the predicted spectral drift value of each sensing area into the corresponding predicted value of the measured parameter.

[0009] Furthermore, in step 1, according to the physical characteristics of the Rayleigh scattering signal in optical fiber, the steps for generating multiple training sample data with long-distance transmission degradation characteristics in batches using data synthesis are as follows: Step 11: Construct a frequency domain Rayleigh scattering spectrum that conforms to the inherent characteristics of fiber Rayleigh scattering; Step 12: Perform frequency domain attenuation and asymmetric degradation modeling on the frequency domain Rayleigh scattering spectrum to simulate the frequency selective attenuation and independent degradation components caused by long-distance fiber transmission, and form a frequency domain reference signal and a frequency domain measurement signal. Step 13: Convert the frequency domain reference signal and frequency domain measurement signal into a range domain reference signal and a range domain measurement signal, and apply a random spectral shift to the range domain measurement signal; Step 14: Based on the signal length of a single sensing region on the sensing fiber, extract the signals at the same position in the range domain reference signal and the range domain measurement signal after applying drift to obtain the local range domain reference signal and the local range domain measurement signal. Step 15: Superimpose typical disturbance components of a long-distance OFDR system onto the local range domain reference signal and the local range domain measurement signal to simulate the signal distortion, noise, and amplitude variation of the sensing fiber in a simulated real measurement environment, thereby obtaining the local reference simulation signal and the local measurement simulation signal. Step 16: Perform amplitude normalization processing and dual-channel integration on the local reference analog signal and the local measurement analog signal, and use the applied spectral drift amount as the corresponding spectral drift label to form a single training sample data; Step 17: Repeat steps 11-16 to obtain a sufficient number of training sample data.

[0010] Furthermore, in step 11, the steps for constructing the frequency domain Rayleigh scattering spectrum that conforms to the inherent characteristics of fiber Rayleigh scattering are as follows: Step 111: Determine the signal redundancy length of the frequency domain Rayleigh scattering spectrum based on the maximum spectral shift of the sensing fiber. The signal redundancy length is as follows:

[0011] in, The maximum spectral shift is given by , where 'a' is a redundancy constant for spectral shift truncation, used to ensure signal integrity after truncation. This is for rounding up; Step 112: Based on the signal length of a single sensing region on the sensing fiber, and in conjunction with the signal redundancy length, determine the original signal length of the frequency domain Rayleigh scattering spectrum. The length of the original signal is as follows:

[0012] Where L is the signal length of a single sensing area; Step 113: Generate two sets of Gaussian random sequences that have the original signal length and are mutually independent; Step 114: Based on two sets of Gaussian random sequences, construct a complex form of random Rayleigh scattering signal. The random Rayleigh scattering signal is shown below:

[0013] in, and There are two sets of Gaussian random sequences, corresponding to the in-phase and quadrature components of the light field, respectively; j is the imaginary unit; Step 115: Perform a Fourier transform on the random Rayleigh scattering signal to transform the random Rayleigh scattering signal from the time domain to the frequency domain, and obtain the frequency domain Rayleigh scattering spectrum.

[0014] Furthermore, in step 12, the steps for performing frequency domain attenuation and asymmetric degradation modeling on the frequency domain Rayleigh scattering spectrum to simulate the frequency-selective attenuation and independent degradation components caused by long-distance fiber transmission are as follows: Step 121: Apply an exponential frequency-domain attenuation factor to the frequency-domain Rayleigh scattering spectrum to simulate the differential attenuation of different frequency components under long-distance fiber transmission, thereby obtaining the attenuated Rayleigh scattering signal. The exponential frequency domain attenuation factor is as follows:

[0015] Where f is the frequency variable and k is the random decay coefficient. It is a natural exponential function; The attenuated Rayleigh scattering signal is shown below:

[0016] in, The frequency domain Rayleigh scattering spectrum; Step 122: Generate two sets of decorrelated random spectra with the same signal length as the attenuated Rayleigh scattering signal and independent of each other. Then, linearly mix these decorrelated random spectra with the attenuated Rayleigh scattering signal using different decorrelated random spectra to obtain the frequency domain reference signal and the frequency domain measurement signal. The frequency domain reference signal is shown below:

[0017] The frequency domain measurement signal is shown below:

[0018] in, and These are the decorrelated random spectrum and degradation coefficients of the frequency domain reference signal, respectively. and These are the decorrelated random spectrum and degradation coefficient of the frequency domain measurement signal, respectively; and satisfy the following conditions: > This is to keep the frequency domain reference signal relatively stable, while simulating the coherence degradation characteristics of the frequency domain measurement signal under long-distance transmission in optical fiber.

[0019] Furthermore, in step 13, the steps of converting the frequency domain reference signal and the frequency domain measurement signal into a range domain reference signal and a range domain measurement signal, and applying a random spectral shift to the range domain measurement signal are as follows: Step 131: Perform inverse Fourier transform on the frequency domain reference signal and the frequency domain measurement signal respectively, and obtain the signal envelope amplitude to obtain the range domain reference signal and the range domain measurement signal; Step 132: Randomly generate a spectral drift within the maximum spectral drift range of the sensing fiber, and apply the randomly generated spectral drift to the distance domain measurement signal to obtain the distance domain measurement signal after drift application.

[0020] Furthermore, in step 14, the local distance domain reference signal is as follows:

[0021] The local range domain measurement signal is shown below:

[0022] in, The range domain reference signal, For the range domain measurement signal after applying the drift, L is the signal length of a single sensing area, n is the sampling point number in the range domain reference signal and the range domain measurement signal, n=0,1,2,...,L-1; P is the signal redundancy length of the range domain reference signal and the range domain measurement signal, and s is the randomly generated spectral drift amount.

[0023] Furthermore, the typical perturbation components of the long-distance OFDR include random gain, slowly varying common-mode amplitude envelope, local signal fading, baseline drift, multiplicative envelope artifacts, and asymmetric random noise. In step 15, the typical perturbation components of the long-distance OFDR system are superimposed on the local range domain reference signal and the local range domain measurement signal to simulate the signal distortion, noise, and amplitude variations of the sensing fiber in a simulated real measurement environment. The steps are as follows: Step 151: Set corresponding random gain coefficients for the local range domain reference signal and the local range domain measurement signal respectively, and perform amplitude scaling on the local range domain reference signal and the local range domain measurement signal to simulate the light intensity and gain fluctuations of the OFDR system. The local range domain reference signal after random gain is shown below:

[0024] The local range domain measurement signal after random gain is shown below:

[0025] in, and These are the random gain coefficients of the local range domain reference signal and the local range domain measurement signal, respectively. and These are the local range domain reference signal and the local range domain measurement signal before random gain, respectively. Step 152: Construct a low-frequency sinusoidal envelope signal and synchronously apply the envelope signal to the local range domain reference signal and the local range domain measurement signal after random gain to simulate the overall signal amplitude fluctuation along the fiber optic distance. The envelope signal is shown below:

[0026] Where A is the amplitude coefficient, f is the envelope frequency, and φ is the initial phase. The signal length of a single sensing area; The local range domain reference signal after amplitude envelope processing is shown below:

[0027] The local range domain measurement signal after amplitude envelope processing is shown below: ; Step 153: Set the fading start position, fading length, and fading coefficient for local signal fading. Within the fading interval, locally compress the range-domain reference signal and the range-domain measurement signal to simulate local signal fading at the far end of the optical fiber. The fading intervals are as follows:

[0028] in, This is the starting position of the fading. The fading length; The locally compressed local range domain reference signal is shown below:

[0029] The local range domain measurement signal after local compression is shown below:

[0030] Wherein, β is the fading coefficient; Step 154: Construct a baseline drift term and add it to the locally compressed local range domain reference signal and local range domain measurement signal to address system drift, de-skewing error processing, and low-frequency baseline drift caused by environmental changes. The local range domain reference signal with baseline drift added is shown below:

[0031] The local range domain measurement signal with baseline drift added is shown below:

[0032] in, For the baseline drift term; Step 155: Construct a multiplicative envelope artifact signal and add the multiplicative envelope artifact signal to the local range domain measurement signal to simulate multiplicative disturbances caused by local amplitude modulation, system fluctuations, and filter residues. The multiplicative envelope artifact signal is shown below:

[0033] Where q is the artifact amplitude coefficient. For artifact frequency, is the initial phase of the artifact, and L is the signal length of a single sensing region; The local range domain measurement signal processed by the multiplicative envelope is shown below:

[0034] Step 156: Construct two sets of random noise signals with the same signal length as the local range domain reference signal and the local range domain measurement signal, and add the two sets of random noise signals to the local range domain reference signal and the local range domain measurement signal respectively, to finally obtain the local reference analog signal and the local measurement analog signal. The local reference analog signal is shown below:

[0035] The local measurement analog signal is shown below:

[0036] in, The random noise signal of the local reference analog signal. The random noise signal of the local measurement analog signal, and satisfying the following conditions: < .

[0037] Furthermore, in step 16, the steps of performing amplitude normalization and dual-channel integration on the local reference analog signal and the local measurement analog signal, and using the applied spectral shift amount as the corresponding spectral shift label to form a single training sample data are as follows: Step 161: Perform mean removal and standardization processing on the local reference analog signal and the local measurement analog signal respectively to obtain a standardized reference analog signal and a standardized measurement analog signal with the same amplitude scale. The standardized reference analog signal is shown below:

[0038] The standardized measurement analog signal is shown below:

[0039] in, The local reference analog signal The mean, The local reference analog signal Standard deviation; For the local reference measurement signal The mean, The local reference analog signal The standard deviation; ε is a local constant used to avoid the denominator being zero; Step 162: Stack the standardized reference analog signal and the standardized measurement analog signal in dual channels to obtain a two-dimensional training matrix. The two-dimensional training matrix is ​​shown below:

[0040] in, The standardized reference analog signal, The standardized measurement analog signal; Step 163: Normalize the spectral shift to obtain the spectral shift label of the two-dimensional training matrix. The spectral drift tag is shown below:

[0041] in, The spectral shift amount, This is the maximum spectral shift.

[0042] Furthermore, in step 4, the steps for constructing the corresponding inference sample data based on the local reference real signal and the local measurement real signal of each sensing area are as follows: Step 41: Perform mean removal and standardization processing on the local reference real signals and local measurement real signals of each sensing area to obtain multiple sets of standardized reference real signals and standardized measurement real signals with the same amplitude scale. The standardized real reference signal for the i-th sensing region is shown below:

[0043] The standardized measurement data of the i-th sensing region is shown below:

[0044] in, For the local reference real signal of the i-th sensing region The mean, For the local reference real signal of the i-th sensing region Standard deviation; For the local measurement real signal of the i-th sensing region The mean, For the local measurement real signal of the i-th sensing region The standard deviation; ε is a local constant used to avoid the denominator being zero; Step 42: Stack the standardized reference real signal and the standardized measurement real signal of each sensing region in a dual-channel configuration to obtain a two-dimensional inference matrix for each sensing region. The two-dimensional inference matrix of the i-th sensing region is shown below:

[0045] in, For the standardized reference real signal of the i-th sensing region, This represents the standardized measurement real signal for the i-th sensing region.

[0046] A long-range OFDR fast demodulation device based on synthetic data training neural network includes a processor and a memory connected together. The memory stores a computer program and a neural network model for the processor to execute. When the processor executes the computer program, it calls the neural network model to perform the above-mentioned long-range OFDR fast demodulation method.

[0047] The present invention has the following beneficial effects: This invention abandons the traditional method of relying on on-site collection and manual annotation of measured samples. Instead, it uses a full-process algorithm to synthesize training samples. By simulating the entire physical process of fiber Rayleigh scattering signal generation, frequency domain attenuation, coherence degradation, and various types of signal disturbances, it can automatically and in large quantities generate training samples without occupying on-site testing equipment and working conditions. This significantly shortens the sample preparation cycle and reduces R&D and application costs.

[0048] This invention synthesizes samples that closely match the characteristics of real-world signals by sequentially simulating frequency-selective attenuation, dual-channel asymmetric coherent degradation, and various typical disturbances such as random gain, local signal fading, baseline drift, and asymmetric noise. During data synthesis, attenuation coefficients, degradation coefficients, spectral drift, and various disturbance parameters are randomly assigned, generating a massive number of samples covering the entire range and multiple operating conditions. This effectively enriches sample diversity and ensures the model has good detection capabilities in different regions at both the near and far ends of the optical fiber. Simultaneously, independent decorrelation spectra and differentiated degradation and noise configurations are used for the reference and measurement signals to accurately reproduce the inherent differences between the two signals, conforming to the actual signal characteristics of OFDR. In the signal truncation stage, a redundant length is set based on the maximum spectral drift, and the discrete sampling rules are adapted through rounding up, effectively avoiding signal truncation and data failure under extreme drift conditions, ensuring the integrity and validity of each synthesized sample. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the steps of the long-distance OFDR fast demodulation method provided by the present invention.

[0050] Figure 2 This is a step-by-step flowchart of step 1 in the long-distance OFDR fast demodulation method provided by the present invention.

[0051] Figure 3 This is a flowchart illustrating the steps of step 11 in the long-distance OFDR fast demodulation method provided by the present invention.

[0052] Figure 4This is a step-by-step flowchart of step 12 in the long-distance OFDR fast demodulation method provided by the present invention.

[0053] Figure 5 This is a flowchart illustrating the steps of step 13 in the long-distance OFDR fast demodulation method provided by the present invention.

[0054] Figure 6 This is a step-by-step flowchart of step 15 in the long-distance OFDR fast demodulation method provided by the present invention.

[0055] Figure 7 This is a step-by-step flowchart of step 16 in the long-distance OFDR fast demodulation method provided by the present invention.

[0056] Figure 8 This is a flowchart illustrating step 4 of the long-distance OFDR fast demodulation method provided by the present invention.

[0057] Figure 9 This is a schematic diagram of the architecture of a one-dimensional convolutional neural network in the long-distance OFDR fast demodulation method provided by the present invention.

[0058] Figure 10 The training and validation loss convergence diagram of the neural network model in the long-distance OFDR fast demodulation method provided by this invention.

[0059] Figure 11 The deviation distribution diagram between the predicted drift and the actual drift of the neural network model in the long-distance OFDR fast demodulation method provided by the present invention.

[0060] Figure 12 This is a spectral drift diagram predicted by the model in the long-distance OFDR fast demodulation method provided by the present invention. Detailed Implementation

[0061] The present invention will now be described in detail with reference to the accompanying drawings and embodiments, examples of which are shown in the drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0062] In the description of this invention, it should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0063] Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first," "second," or "third" may explicitly or implicitly include one or more of that feature. In the description of this invention, "multiple" means two or more, unless otherwise explicitly specified.

[0064] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," "fixing," and "setting," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0065] Example 1 like Figure 1 As shown, a fast demodulation method for long-range OFDR based on synthetic data-trained neural networks includes the following steps: Step 1: Based on the physical characteristics of fiber Rayleigh scattering signals, multiple training sample data with long-distance transmission degradation characteristics are generated in batches using a data synthesis method. Each training sample data includes a local reference analog signal, a local measurement analog signal and its corresponding spectral drift label.

[0066] The purpose of step 1 is to generate a large number of training samples with rich features that closely match the real-world conditions of long-distance OFDR through algorithm simulation, so as to replace the high-cost and difficult-to-collect experimental labeled samples and provide sufficient data sources for supervised training of neural networks.

[0067] Specifically, such as Figure 2 As shown, in step 1, according to the physical characteristics of fiber Rayleigh scattering signals, the steps for generating multiple training sample data with long-distance transmission degradation characteristics in batches using data synthesis are as follows: Step 11: Construct a frequency domain Rayleigh scattering spectrum that conforms to the inherent characteristics of fiber Rayleigh scattering.

[0068] The purpose of step 11 is to generate a basic signal that conforms to physical laws by simulating the original Rayleigh scattering light field characteristics of the optical fiber under test, which serves as the original carrier for subsequent frequency domain processing and signal degradation.

[0069] Specifically, such as Figure 3 As shown, in step 11, the steps for constructing the frequency domain Rayleigh scattering spectrum that conforms to the inherent characteristics of fiber Rayleigh scattering are as follows: Step 111: Determine the signal redundancy length of the frequency domain Rayleigh scattering spectrum based on the maximum spectral drift of the sensing fiber.

[0070] In step 111, the maximum spectral shift is the maximum value of the spectral shift of the sensing optical fiber under the influence of the measured parameter, which corresponds to the measurement range of the measured parameter.

[0071] The signal redundancy length is as follows:

[0072] in, The maximum spectral shift is given by , where 'a' is a redundancy constant for spectral shift truncation, used to ensure signal integrity after truncation. This is for rounding up.

[0073] The purpose of step 111 is to reserve a redundant signal of length P at each end of the effective sensing signal to accommodate full-range spectral drift and prevent index out-of-bounds errors and loss of effective waveforms during subsequent signal truncation. In this embodiment, the redundancy constant a is set to 5.

[0074] Step 112: Determine the original signal length of the frequency domain Rayleigh scattering spectrum based on the signal length of a single sensing region on the sensing fiber and the signal redundancy length.

[0075] In step 112, the single sensing region refers to an independent physical detection area on the sensing optical fiber, corresponding to the signal range of a single input to the neural network model, and is also the smallest unit for finally calculating the measured parameter. The signal length of a single sensing region refers to the total number of sampling points in a single sensing region, which is the fixed input length of the neural network model.

[0076] The signal length of a single sensing area is set according to the hardware sampling rate and spatial resolution. In this embodiment, the value is 1024 sampling points; the maximum spectral drift is determined based on the full-range calibration result of the measured parameter, and the value range is [100, 300] sampling points.

[0077] The length of the original signal is as follows:

[0078] Where L is the signal length of a single sensing area.

[0079] Step 113: Generate two sets of Gaussian random sequences that have the original signal length and are independent of each other.

[0080] In step 113, since the two orthogonal components of the Rayleigh scattering light field follow independent Gaussian distributions, the inherent statistical characteristics of the light field can be reproduced through the Gaussian random sequence. Based on the Gaussian distribution random number algorithm, two sets of discrete random sequences with a mean of 0 and configurable variance are generated. The two sets of sequences are completely independent of each other and have no data correlation. The variance of the Gaussian sequence ranges from [0.8 to 1.2] to ensure that the signal amplitude is within a reasonable range.

[0081] Step 114: Construct a complex form of random Rayleigh scattering signal based on two sets of Gaussian random sequences.

[0082] In step 114, the random Rayleigh scattering signal is as follows:

[0083] in, and There are two sets of Gaussian random sequences, corresponding to the in-phase and quadrature components of the light field, respectively; j is the imaginary unit.

[0084] Step 115: Perform a Fourier transform on the random Rayleigh scattering signal to transform the random Rayleigh scattering signal from the time domain to the frequency domain, and obtain the frequency domain Rayleigh scattering spectrum.

[0085] The purpose of step 115 is to transform the random Rayleigh scattering signal from the time domain to the frequency domain, so as to facilitate subsequent simulation of long-distance frequency-selective attenuation and frequency domain coherent degradation of optical fibers.

[0086] Step 12: Perform frequency domain attenuation and asymmetric degradation modeling on the frequency domain Rayleigh scattering spectrum to simulate the frequency selective attenuation and independent degradation components caused by long-distance fiber transmission, forming a frequency domain reference signal and a frequency domain measurement signal.

[0087] The purpose of step 12 is to reproduce the two core signal degradation characteristics of long-distance OFDR: first, the frequency domain attenuation characteristics of optical signals of different frequencies with different transmission losses in optical fibers; and second, the asymmetric coherent degradation characteristics between the reference signal and the measurement signal, thus restoring the essential differences in signals under long-distance operating conditions.

[0088] Specifically, such as Figure 4 As shown, in step 12, the steps for performing frequency domain attenuation and asymmetric degradation modeling on the frequency domain Rayleigh scattering spectrum to simulate the frequency-selective attenuation and independent degradation components caused by long-distance fiber transmission are as follows: Step 121: Apply an exponential frequency-domain attenuation factor to the frequency-domain Rayleigh scattering spectrum to simulate the differential attenuation of different frequency components under long-distance optical fiber transmission, thereby obtaining an attenuated Rayleigh scattering signal.

[0089] In step 121, the exponential frequency domain attenuation factor is as follows:

[0090] Where f is the frequency variable and k is the random decay coefficient. It is a natural exponential function; The attenuated Rayleigh scattering signal is shown below:

[0091] in, The frequency domain Rayleigh scattering spectrum is described.

[0092] The random attenuation coefficient k is used to simulate the differentiated attenuation degree of different transmission distances and different optical fiber links, and its value range is [0.001, 0.01]. In long-distance optical fiber transmission, the loss of high-frequency signals is greater than that of low-frequency signals, and the exponential attenuation function is used to simulate this physical law. The larger the attenuation coefficient k is, the more obvious the amplitude attenuation of the higher frequency signal is.

[0093] Step 122: Generate two sets of decorrelated random spectra with the same signal length as the attenuated Rayleigh scattering signal and independent of each other, and use different decorrelated random spectra to linearly mix with the attenuated Rayleigh scattering signal to obtain the frequency domain reference signal and the frequency domain measurement signal.

[0094] The purpose of step 122 is to simulate the signal coherence loss caused by long-distance descrambling filtering and laser phase noise accumulation; to configure independent spectra and different degradation coefficients for the two channels, to achieve asymmetric degradation, and to fit the actual working condition of "stable reference signal and severely degraded measurement signal".

[0095] The frequency domain reference signal is shown below:

[0096] The frequency domain measurement signal is shown below:

[0097] in, and These are the decorrelated random spectrum and degradation coefficients of the frequency domain reference signal, respectively. and These are the decorrelated random spectrum and degradation coefficient of the frequency domain measurement signal, respectively; and satisfy the following conditions: > This is to keep the frequency domain reference signal relatively stable, while simulating the coherence degradation characteristics of the frequency domain measurement signal under long-distance transmission in optical fiber.

[0098] and Two independent frequency-domain random sequences are used to simulate the decorrelation spectrum corresponding to system noise and coherence loss. The original attenuated signal and the decorrelation spectrum are weighted and mixed by a degradation coefficient; the larger the degradation coefficient, the more severe the signal coherence loss. The degradation coefficient of the frequency-domain reference signal ranges from [0.02, 0.1], and the degradation coefficient of the frequency-domain measurement signal... The value range is [0.1, 0.4], and it always satisfies... > This ensures that the degradation of the measured signal is higher than that of the reference signal.

[0099] Step 13: Convert the frequency domain reference signal and frequency domain measurement signal into a range domain reference signal and a range domain measurement signal, and apply a random spectral shift to the range domain measurement signal.

[0100] The purpose of step 13 is to restore the frequency domain signal to a distance domain signal that can be processed by engineering algorithms and models, to artificially apply a controllable spectral drift, and to extract a standard length effective window, thereby simulating the signal drift caused by the change of the measured parameter, and generating the prerequisites for the drift truth.

[0101] Specifically, such as Figure 5 As shown, in step 13, the steps of converting the frequency domain reference signal and the frequency domain measurement signal into a range domain reference signal and a range domain measurement signal, and applying a random spectral shift to the range domain measurement signal are as follows: Step 131: Perform inverse Fourier transform on the frequency domain reference signal and the frequency domain measurement signal respectively, and obtain the signal envelope amplitude to obtain the range domain reference signal and the range domain measurement signal.

[0102] The purpose of step 131 is to complete the inverse transformation of the signal from the frequency domain to the range domain, extract the light intensity envelope signal, and match the signal form actually acquired by the OFDR system.

[0103] Step 132: Randomly generate a spectral drift within the maximum spectral drift range of the sensing fiber, and apply the randomly generated spectral drift to the distance domain measurement signal to obtain the distance domain measurement signal after drift application.

[0104] The purpose of step 131 is to simulate the Rayleigh scattering spectral drift caused by changes in the measured parameters such as temperature and strain, and to randomly assign values ​​to ensure that the sample covers the entire drift range.

[0105] The spectral shift is a positive sampling point, with a value range of [-]. , The positive and negative signs correspond to the spectral drift direction, covering conditions such as heating / cooling, stretching / compression, etc.

[0106] Step 14: Based on the signal length of a single sensing region on the sensing fiber, extract the signals at the same position in the range domain reference signal and the range domain measurement signal after applying drift to obtain the local range domain reference signal and the local range domain measurement signal.

[0107] The purpose of step 14 is to extract local range domain reference signals and local range domain measurement signals with the signal length of a single sensing region from the same position of the range domain reference signal and the range domain measurement signal after applying drift, respectively, so as to simulate the sliding window value during OFDR signal demodulation.

[0108] The local range domain reference signal is shown below:

[0109] The local range domain measurement signal is shown below:

[0110] in, The range domain reference signal, For the range domain measurement signal after applying the drift, L is the signal length of a single sensing area, n is the sampling point number in the range domain reference signal and the range domain measurement signal, n=0,1,2,...,L-1; P is the signal redundancy length of the range domain reference signal and the range domain measurement signal, and s is the randomly generated spectral drift amount.

[0111] Step 15: Superimpose typical disturbance components of a long-distance OFDR system onto the local range domain reference signal and the local range domain measurement signal to simulate the signal distortion, noise, and amplitude changes of the sensing fiber in a simulated real measurement environment, thereby obtaining the local reference simulation signal and the local measurement simulation signal.

[0112] The purpose of step 15 is to layer and superimpose all typical signal disturbances in the long-distance OFDR field, comprehensively reproduce the real working conditions such as signal distortion, amplitude fluctuation, and noise interference at the far end, and further reduce the gap between the analog signal and the measured sample.

[0113] In this example, the typical perturbation components of long-distance OFDR include random gain, slowly varying common-mode amplitude envelope, local signal fading, baseline drift, multiplicative envelope artifacts, and asymmetric random noise.

[0114] Specifically, such as Figure 6 As shown, in step 15, the typical disturbance components of a long-distance OFDR system are superimposed on the local range domain reference signal and the local range domain measurement signal to simulate the signal distortion, noise, and amplitude variation of the sensing fiber in a simulated real measurement environment. Step 151: Set corresponding random gain coefficients for the local range domain reference signal and the local range domain measurement signal respectively, and perform amplitude scaling on the local range domain reference signal and the local range domain measurement signal to simulate the light intensity and gain fluctuations of the OFDR system.

[0115] In step 151, the laser and optical path devices of the OFDR system exhibit light intensity fluctuations. This phenomenon is simulated by scaling the overall signal amplitude using a random gain coefficient.

[0116] The local range domain reference signal after random gain is shown below:

[0117] The local range domain measurement signal after random gain is shown below:

[0118] in, and These are the random gain coefficients of the local range domain reference signal and the local range domain measurement signal, respectively. and These are the local range domain reference signal and the local range domain measurement signal before random gain, respectively.

[0119] The random gain coefficient and The value range is [0.7, 1.3].

[0120] Step 152: Construct a low-frequency sinusoidal envelope signal and apply the envelope signal synchronously to the local distance domain reference signal and the local distance domain measurement signal after random gain, so as to simulate the overall signal amplitude fluctuation along the fiber distance.

[0121] The purpose of step 152 is to use a low-frequency sinusoidal signal to simulate the slow amplitude fluctuations caused by long-distance optical fiber transmission, with the two signals sharing the same envelope to reflect common-mode distortion characteristics.

[0122] The envelope signal is shown below:

[0123] Where A is the amplitude coefficient, f is the envelope frequency, and φ is the initial phase. The signal length of a single sensing area.

[0124] The amplitude coefficient A ranges from [0.1, 0.3], the envelope frequency f ranges from [1, 5] (low-frequency period within a single window), and the initial phase... The value of is in the range of [0, 2π], and is randomly selected.

[0125] The local range domain reference signal after amplitude envelope processing is shown below:

[0126] The local range domain measurement signal after amplitude envelope processing is shown below: .

[0127] Step 153: Set the fading start position, fading length, and fading coefficient for local signal fading, and perform local compression on the range domain reference signal and range domain measurement signal within the fading interval to simulate local signal fading at the far end of the optical fiber.

[0128] In step 153, local defects, bends, and far-end losses in the optical fiber can cause a sudden drop in the local signal amplitude. After defining the interval, the amplitude is compressed by the fading coefficient to simulate local fading.

[0129] The fading intervals are as follows:

[0130] in, This is the starting position of the fading. The fading length is denoted as .

[0131] The fading start position The range of values ​​for is [0, ...]. ], randomly selected, the fading length The value range is [100, 400] sampling points.

[0132] The locally compressed local range domain reference signal is shown below:

[0133] The local range domain measurement signal after local compression is shown below:

[0134] Wherein, β is the fading coefficient, and its value ranges from [0.4, 0.8]. The smaller the coefficient, the more severe the fading.

[0135] Step 154: Construct a baseline drift term and add the baseline drift term to the locally compressed local range domain reference signal and the local range domain measurement signal to address system drift, deskewing error processing, and low-frequency baseline drift caused by environmental changes.

[0136] In step 154, circuit temperature drift and device bias will cause signal baseline drift. The drift amplitude of the measured signal is greater than that of the reference signal. The baseline drift term generated by a linear / low-frequency slow-varying function can be used for simulation.

[0137] The local range domain reference signal with baseline drift added is shown below:

[0138] The local range domain measurement signal with baseline drift added is shown below:

[0139] in, This refers to the baseline drift term.

[0140] Step 155: Construct a multiplicative envelope artifact signal and add the multiplicative envelope artifact signal to the local range domain measurement signal to simulate multiplicative disturbances caused by local amplitude modulation, system fluctuations and filter residues.

[0141] In step 155, long-distance measurement links are prone to periodic artifact interference. A sinusoidal multiplicative signal is used to simulate this interference, which is added only to the measurement signal to match the actual working conditions.

[0142] The multiplicative envelope artifact signal is shown below:

[0143] Where q is the artifact amplitude coefficient, with a value range of [0.05, 0.2]. The value of is the artifact frequency, which ranges from [3, 8]. The initial phase of the artifact is defined as [0, 2π], and is randomly selected.

[0144] The local range domain measurement signal processed by the multiplicative envelope is shown below: .

[0145] Step 156: Construct two sets of random noise signals with the same signal length as the local range domain reference signal and the local range domain measurement signal, and add the two sets of random noise signals to the local range domain reference signal and the local range domain measurement signal respectively, so as to finally obtain the local reference simulation signal and the local measurement simulation signal.

[0146] In step 156, the reference signal link is short and has low noise, while the measurement signal has greater noise over long distances, generating two sets of independent Gaussian random noise to achieve asymmetric noise addition.

[0147] The local reference analog signal is shown below:

[0148] The local measurement analog signal is shown below:

[0149] in, The random noise signal of the local reference analog signal. The random noise signal of the local measurement analog signal, and satisfying the following conditions: < .

[0150] Step 15: Based on the signal length of a single sensing region on the sensing fiber, extract the signals at the same position in the range domain reference signal and the range domain measurement signal after applying drift to obtain the local range domain reference signal and the local range domain measurement signal.

[0151] Step 16: Perform amplitude normalization processing and dual-channel integration on the local reference analog signal and the local measurement analog signal, and use the applied spectral drift amount as the corresponding spectral drift label to form a single training sample data.

[0152] In step 16, the two signals are integrated into the dual-channel format required by the neural network model, and the amplitude difference is eliminated by standardization; the spectral drift is normalized to generate training labels, and a single training sample is created.

[0153] Specifically, such as Figure 7 As shown, in step 16, the local reference analog signal and the local measurement analog signal are subjected to amplitude normalization and dual-channel integration, and the applied spectral shift is used as the corresponding spectral shift label to form a single training sample data. The steps are as follows: Step 161: Perform mean removal and standardization processing on the local reference analog signal and the local measurement analog signal respectively to obtain a standardized reference analog signal and a standardized measurement analog signal with the same amplitude scale.

[0154] The purpose of step 161 is to unify the amplitude distribution of all samples to avoid amplitude differences affecting model training; and to add a very small constant to prevent the denominator from being zero.

[0155] The standardized reference analog signal is shown below:

[0156] The standardized measurement analog signal is shown below:

[0157] in, The local reference analog signal The mean, The local reference analog signal Standard deviation; For the local reference measurement signal The mean, The local reference analog signal The standard deviation; ε is a minimal constant used to avoid the denominator being zero.

[0158] In this embodiment, the value of the minimum constant ε is 10. -8 .

[0159] Step 162: Stack the standardized reference analog signal and the standardized measurement analog signal in two channels to obtain a two-dimensional training matrix.

[0160] The purpose of step 162 is to combine the standardized reference analog signal and the standardized measurement analog signal into a dual-channel input format adapted to the neural network model.

[0161] The two-dimensional training matrix is ​​shown below:

[0162] in, The standardized reference analog signal, The standardized measurement analog signal is referred to here.

[0163] Step 163: Normalize the spectral drift to obtain the spectral drift label of the two-dimensional training matrix.

[0164] The purpose of step 163 is to normalize the large range of spectral shifts to the interval [-1, 1] to fit the output range of the regression model.

[0165] The spectral drift tag is shown below:

[0166] in, The spectral shift amount, This is the maximum spectral shift.

[0167] Step 17: Repeat steps 11-16 to obtain a sufficient number of training sample data.

[0168] In step 17, a large number of diverse samples are generated in a loop. In this embodiment, the total number of synthesized samples is no less than 20,000 to ensure the training effect of the model.

[0169] Step 2: Use the training sample data to perform supervised training on the preset neural network model to obtain a drift prediction model that has been trained.

[0170] In step 2, all training sample data are randomly divided into a training dataset and a validation dataset. The training dataset accounts for 70% to 80% of the total training samples, and the validation dataset accounts for 20% to 30% of the total training samples. The neural network model is then trained using the training dataset to update the network weights of the neural network model through the backpropagation algorithm. The validation dataset is then used to verify the test results of the neural network model. If the verification is successful, the training of the neural network model is completed.

[0171] During supervised training, the loss function used by the neural network model is as follows:

[0172] in, To simulate the generated spectral drift label values, is the predicted value of the spectral drift output by the neural network model, and N is the number of samples in a single batch of data during supervised training.

[0173] In this embodiment, the neural network model is a one-dimensional convolutional neural network, such as... Figure 9 As shown, it consists of an input layer, a multi-level one-dimensional convolutional layer, a max pooling layer, a global average pooling layer, and a fully connected regression layer. Its training process uses the Adam optimization algorithm.

[0174] The input layer of the one-dimensional convolutional neural network uses a dual-channel input consisting of a reference signal and a measurement signal. Its input size is 2×1×1024, where the two input channels correspond to the reference signal and the measurement signal, respectively, and the signal length of each input channel is 1024 sampling points.

[0175] The front end of the one-dimensional convolutional neural network consists of multiple one-dimensional convolutional layers. Each one-dimensional convolutional layer extracts the local morphological differences and displacement-related features between the reference signal and the measurement signal through one-dimensional convolutional kernels of different scales, and combines pooling operations to gradually compress the time dimension and reduce the subsequent computational load.

[0176] In this embodiment, the one-dimensional convolutional neural network includes a first one-dimensional convolutional layer, a second one-dimensional convolutional layer, and a third one-dimensional convolutional layer. The input reference signal and measurement signal are passed through the first one-dimensional convolutional layer to obtain a 32×1×1024 output feature map. Then, they are downsampled to a 32×1×512 output feature map by the first max pooling layer. After passing through the second one-dimensional convolutional layer, a 64×1×512 output feature map is obtained. It is further downsampled to 64×1×256 by the second max pooling layer. The third one-dimensional convolutional layer continues to extract higher-level output feature maps, finally forming a deep feature output of 128 / 256×1×128.

[0177] After completing the convolutional feature extraction, the one-dimensional convolutional neural network uses global average pooling to compress the temporal features into a feature vector of length 256. This process reduces the number of parameters in the fully connected regression layer, lowers the model inference overhead, and preserves spectral drift information within the window. Finally, the feature vector is input into multiple fully connected regression layers, undergoing a step-by-step mapping of 512-256-128-1, ultimately outputting a normalized spectral drift prediction value.

[0178] Figure 10 The graph shows the convergence of the training and validation loss of the neural network model. As can be seen from the graph, the prediction error curves of the training dataset and the validation dataset both show a continuous and rapid decline before stabilizing, and the gap between them continues to narrow. There are no violent fluctuations in the later stages, indicating that the model has fully learned the spectral features, can accurately predict the spectral drift, has good generalization ability, and has low error on unfamiliar validation spectra. There is no overfitting or underfitting, and training can be stopped.

[0179] Figure 11 The graph shows the distribution of the deviation between the predicted drift and the actual drift when the neural network model is trained using the validation dataset. As can be seen from the graph, the predicted drift (red dots) within ±5% accounts for 90.14%, which has engineering value. The prediction accuracy can be continuously improved by increasing the training sample data. The predicted drift (blue dots) outside ±5% is uniformly distributed, indicating that the network performance meets expectations and there is no systematic error.

[0180] Step 3: Acquire the local reference real signal and local measurement real signal of each sensing area on the sensing fiber using an optical frequency domain reflectometer.

[0181] In step 3, the optical frequency domain reflectometer first collects Rayleigh scattering signals when the sensing fiber is in the reference state and the measurement state to obtain the true reference signal and the true measurement signal of the entire sensing fiber. Then, it performs a fast Fourier transform on the true reference signal and the true measurement signal to transform them from the frequency domain to the distance domain. Then, according to the signal length of a single sensing region and the position distribution of each sensing region on the sensing fiber, it performs sliding window values ​​on the true reference signal and the true measurement signal to obtain the local reference true signal and the local measurement true signal of each sensing region.

[0182] Step 4: Based on the local reference real signal and local measurement real signal of each sensing region, construct corresponding inference sample data, and input the inference sample data of each sensing region into the drift prediction model for forward inference in sequence. Then, receive the spectral drift prediction value of each sensing region output by the drift prediction model in sequence.

[0183] In step 4, the real field signals are preprocessed according to the same rules as the synthetic samples to construct inference samples, so as to quickly predict the spectral drift using the trained drift prediction model.

[0184] Specifically, such as Figure 8 As shown, in step 4, the steps for constructing the corresponding inference sample data based on the local reference real signal and the local measurement real signal of each sensing area are as follows: Step 41: Perform mean removal and standardization processing on the local reference real signal and local measurement real signal of each sensing area to obtain multiple sets of standardized reference real signal and standardized measurement real signal with the same amplitude scale.

[0185] In step 41, the normalized reference real signal of the i-th sensing region is as follows:

[0186] The standardized measurement data of the i-th sensing region is shown below:

[0187] in, For the local reference real signal of the i-th sensing region The mean, For the local reference real signal of the i-th sensing region Standard deviation; For the local measurement real signal of the i-th sensing region The mean, For the local measurement real signal of the i-th sensing region The standard deviation; ε is a local constant used to avoid the denominator being zero; The local minimum constant ε here is consistent with the local minimum constant in the training samples.

[0188] Step 42: Stack the standardized reference real signal and the standardized measurement real signal of each sensing region in two channels to obtain the two-dimensional inference matrix of each sensing region.

[0189] In step 42, the two-dimensional inference matrix of the i-th sensing region is as follows:

[0190] in, For the standardized reference real signal of the i-th sensing region, This represents the standardized measurement real signal for the i-th sensing region.

[0191] Step 5: Based on the inversion function between the pre-calibrated spectral drift and the measured parameter, convert the predicted spectral drift value of each sensing area into the corresponding predicted value of the measured parameter.

[0192] In step 5, the final conversion from spectral drift to physical parameters such as fiber temperature and strain is completed, and the monitoring results usable in the project are output.

[0193] In OFDR systems, the spectral shift is linearly correlated with the measured parameter (such as temperature or strain), which is a common calibration rule in the industry. The inversion function adopts a linear model as shown below:

[0194] Where P is the magnitude of the measured parameter. This is the initial reference value of the measured parameter (the parameter value when the optical fiber is in a reference state). The spectral shift amount, This is the conversion coefficient between the spectral drift and the measured parameter.

[0195] In this embodiment, micro-stresses of 500N and 750N are applied at a position of 970m on the sensing fiber, respectively. The reference and measured real signals from the sensing fiber are input into the drift prediction model, resulting in the spectral drift map predicted by the model, as shown below. Figure 12 As shown in the figure, when the applied micro-stress is 500N, the predicted spectral drift value of the drift prediction model is 495.5, and when the applied micro-stress is 750N, the predicted spectral drift value of the drift prediction model is 755.0. The prediction error is within ±0.1%, indicating that the stress and the amount of spectral drift have a highly linear relationship, and the conversion coefficient k≈1.

[0196] Example 2 A long-distance OFDR fast demodulation device based on synthetic data training neural network includes a processor and a memory connected together. The memory stores a computer program and a neural network model for the processor to execute. When the processor executes the computer program, it calls the neural network model to perform the long-distance OFDR fast demodulation method described in Embodiment 1.

[0197] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention and not to limit them. Although the embodiments of the present invention have been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the embodiments of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fast demodulation method for long-range OFDR based on synthetic data-trained neural networks, characterized in that, Includes the following steps: Step 1: Based on the physical characteristics of fiber Rayleigh scattering signals, multiple training sample data with long-distance transmission degradation characteristics are generated in batches using data synthesis. Each training sample data includes a local reference analog signal, a local measurement analog signal and its corresponding spectral drift label. Step 2: Use the training sample data to perform supervised training on the preset neural network model to obtain a drift prediction model that has completed training; Step 3: Acquire the local reference real signal and local measurement real signal of each sensing area on the sensing fiber using an optical frequency domain reflectometer; Step 4: Based on the local reference real signal and local measurement real signal of each sensing area, construct corresponding inference sample data, and input the inference sample data of each sensing area into the drift prediction model for forward inference in sequence. Then, receive the spectral drift prediction value of each sensing area output by the drift prediction model in sequence. Step 5: Based on the inversion function between the pre-calibrated spectral drift and the measured parameter, convert the predicted spectral drift value of each sensing area into the corresponding predicted value of the measured parameter.

2. The long-distance OFDR fast demodulation method according to claim 1, characterized in that, In step 1, based on the physical characteristics of fiber Rayleigh scattering signals, the steps for generating multiple training sample data with long-distance transmission degradation characteristics in batches using data synthesis are as follows: Step 11: Construct a frequency domain Rayleigh scattering spectrum that conforms to the inherent characteristics of fiber Rayleigh scattering; Step 12: Perform frequency domain attenuation and asymmetric degradation modeling on the frequency domain Rayleigh scattering spectrum to simulate the frequency selective attenuation and independent degradation components caused by long-distance fiber transmission, and form a frequency domain reference signal and a frequency domain measurement signal. Step 13: Convert the frequency domain reference signal and frequency domain measurement signal into a range domain reference signal and a range domain measurement signal, and apply a random spectral shift to the range domain measurement signal; Step 14: Based on the signal length of a single sensing region on the sensing fiber, extract the signals at the same position in the range domain reference signal and the range domain measurement signal after applying drift to obtain the local range domain reference signal and the local range domain measurement signal. Step 15: Superimpose typical disturbance components of a long-distance OFDR system onto the local range domain reference signal and the local range domain measurement signal to simulate the signal distortion, noise, and amplitude variation of the sensing fiber in a simulated real measurement environment, thereby obtaining the local reference simulation signal and the local measurement simulation signal. Step 16: Perform amplitude normalization processing and dual-channel integration on the local reference analog signal and the local measurement analog signal, and use the applied spectral drift amount as the corresponding spectral drift label to form a single training sample data; Step 17: Repeat steps 11-16 to obtain a sufficient number of training sample data.

3. The long-distance OFDR fast demodulation method according to claim 2, characterized in that, In step 11, the steps for constructing the frequency domain Rayleigh scattering spectrum that conforms to the inherent characteristics of fiber Rayleigh scattering are as follows: Step 111: Determine the signal redundancy length of the frequency domain Rayleigh scattering spectrum based on the maximum spectral shift of the sensing fiber. The signal redundancy length is as follows: in, The maximum spectral shift is given by , where 'a' is a redundancy constant for spectral shift truncation, used to ensure signal integrity after truncation. This is for rounding up; Step 112: Based on the signal length of a single sensing region on the sensing fiber, and in conjunction with the signal redundancy length, determine the original signal length of the frequency domain Rayleigh scattering spectrum. The length of the original signal is as follows: Where L is the signal length of a single sensing area; Step 113: Generate two sets of Gaussian random sequences that have the original signal length and are mutually independent; Step 114: Based on two sets of Gaussian random sequences, construct a complex form of random Rayleigh scattering signal. The random Rayleigh scattering signal is shown below: in, and There are two sets of Gaussian random sequences, corresponding to the in-phase and quadrature components of the light field, respectively; j is the imaginary unit; Step 115: Perform a Fourier transform on the random Rayleigh scattering signal to transform the random Rayleigh scattering signal from the time domain to the frequency domain, and obtain the frequency domain Rayleigh scattering spectrum.

4. The long-distance OFDR fast demodulation method according to claim 2, characterized in that, In step 12, the frequency domain attenuation and asymmetric degradation modeling of the frequency domain Rayleigh scattering spectrum is performed to simulate the frequency-selective attenuation and independent degradation components caused by long-distance fiber transmission. The steps are as follows: Step 121: Apply an exponential frequency-domain attenuation factor to the frequency-domain Rayleigh scattering spectrum to simulate the differential attenuation of different frequency components under long-distance fiber transmission, thereby obtaining the attenuated Rayleigh scattering signal. The exponential frequency domain attenuation factor is as follows: Where f is the frequency variable and k is the random decay coefficient. It is a natural exponential function; The attenuated Rayleigh scattering signal is shown below: in, The frequency domain Rayleigh scattering spectrum; Step 122: Generate two sets of decorrelated random spectra with the same signal length as the attenuated Rayleigh scattering signal and independent of each other. Then, linearly mix these decorrelated random spectra with the attenuated Rayleigh scattering signal using different decorrelated random spectra to obtain the frequency domain reference signal and the frequency domain measurement signal. The frequency domain reference signal is shown below: The frequency domain measurement signal is shown below: in, and These are the decorrelated random spectrum and degradation coefficients of the frequency domain reference signal, respectively. and These are the decorrelated random spectrum and degradation coefficient of the frequency domain measurement signal, respectively; and satisfy the following conditions: > This is to keep the frequency domain reference signal relatively stable, while simulating the coherence degradation characteristics of the frequency domain measurement signal under long-distance transmission in optical fiber.

5. The long-distance OFDR fast demodulation method according to claim 2, characterized in that, In step 13, the steps of converting the frequency domain reference signal and the frequency domain measurement signal into a range domain reference signal and a range domain measurement signal, and applying a random spectral shift to the range domain measurement signal are as follows: Step 131: Perform inverse Fourier transform on the frequency domain reference signal and the frequency domain measurement signal respectively, and obtain the signal envelope amplitude to obtain the range domain reference signal and the range domain measurement signal; Step 132: Randomly generate a spectral drift within the maximum spectral drift range of the sensing fiber, and apply the randomly generated spectral drift to the distance domain measurement signal to obtain the distance domain measurement signal after drift application.

6. The long-distance OFDR fast demodulation method according to claim 2, characterized in that, In step 14, the local range domain reference signal is as follows: The local range domain measurement signal is shown below: in, The range domain reference signal, For the range domain measurement signal after applying the drift, L is the signal length of a single sensing area, n is the sampling point number in the range domain reference signal and the range domain measurement signal, n=0,1,2,...,L-1; P is the signal redundancy length of the range domain reference signal and the range domain measurement signal, and s is the randomly generated spectral drift amount.

7. The long-distance OFDR fast demodulation method according to claim 2, characterized in that, The typical perturbation components of the long-distance OFDR include random gain, slowly varying common-mode amplitude envelope, local signal fading, baseline drift, multiplicative envelope artifacts, and asymmetric random noise. In step 15, the typical perturbation components of the long-distance OFDR system are superimposed on the local range domain reference signal and the local range domain measurement signal to simulate the signal distortion, noise, and amplitude variations of the sensing fiber in a simulated real measurement environment. The steps are as follows: Step 151: Set corresponding random gain coefficients for the local range domain reference signal and the local range domain measurement signal respectively, and perform amplitude scaling on the local range domain reference signal and the local range domain measurement signal to simulate the light intensity and gain fluctuations of the OFDR system. The local range domain reference signal after random gain is shown below: The local range domain measurement signal after random gain is shown below: in, and These are the random gain coefficients of the local range domain reference signal and the local range domain measurement signal, respectively. and These are the local range domain reference signal and the local range domain measurement signal before random gain, respectively. Step 152: Construct a low-frequency sinusoidal envelope signal and synchronously apply the envelope signal to the local range domain reference signal and the local range domain measurement signal after random gain to simulate the overall signal amplitude fluctuation along the fiber optic distance. The envelope signal is shown below: Where A is the amplitude coefficient, f is the envelope frequency, and φ is the initial phase. The signal length of a single sensing area; The local range domain reference signal after amplitude envelope processing is shown below: The local range domain measurement signal after amplitude envelope processing is shown below: ； Step 153: Set the fading start position, fading length, and fading coefficient for local signal fading. Within the fading interval, locally compress the range-domain reference signal and the range-domain measurement signal to simulate local signal fading at the far end of the optical fiber. The fading intervals are as follows: in, This is the starting position of the fading. The fading length; The locally compressed local range domain reference signal is shown below: The local range domain measurement signal after local compression is shown below: Wherein, β is the fading coefficient; Step 154: Construct a baseline drift term and add it to the locally compressed local range domain reference signal and local range domain measurement signal to address system drift, de-skewing error processing, and low-frequency baseline drift caused by environmental changes. The local range domain reference signal with baseline drift added is shown below: The local range domain measurement signal with baseline drift added is shown below: in, For the baseline drift term; Step 155: Construct a multiplicative envelope artifact signal and add the multiplicative envelope artifact signal to the local range domain measurement signal to simulate multiplicative disturbances caused by local amplitude modulation, system fluctuations, and filter residues. The multiplicative envelope artifact signal is shown below: Where q is the artifact amplitude coefficient. For artifact frequency, is the initial phase of the artifact, and L is the signal length of a single sensing region; The local range domain measurement signal processed by the multiplicative envelope is shown below: Step 156: Construct two sets of random noise signals with the same signal length as the local range domain reference signal and the local range domain measurement signal, and add the two sets of random noise signals to the local range domain reference signal and the local range domain measurement signal respectively, to finally obtain the local reference analog signal and the local measurement analog signal. The local reference analog signal is shown below: The local measurement analog signal is shown below: in, The random noise signal of the local reference analog signal. The random noise signal of the local measurement analog signal, and satisfying the following conditions: < .

8. The long-distance OFDR fast demodulation method according to claim 1, characterized in that, In step 16, the local reference analog signal and the local measurement analog signal are subjected to amplitude normalization and dual-channel integration, and the applied spectral shift is used as the corresponding spectral shift label to form a single training sample data. The steps are as follows: Step 161: Perform mean removal and standardization processing on the local reference analog signal and the local measurement analog signal respectively to obtain a standardized reference analog signal and a standardized measurement analog signal with the same amplitude scale. The standardized reference analog signal is shown below: The standardized measurement analog signal is shown below: in, The local reference analog signal The mean, The local reference analog signal Standard deviation; For the local reference measurement signal The mean, The local reference analog signal The standard deviation; ε is a local constant used to avoid the denominator being zero; Step 162: Stack the standardized reference analog signal and the standardized measurement analog signal in dual channels to obtain a two-dimensional training matrix. The two-dimensional training matrix is ​​shown below: in, The standardized reference analog signal, The standardized measurement analog signal; Step 163: Normalize the spectral shift to obtain the spectral shift label of the two-dimensional training matrix. The spectral drift tag is shown below: in, The spectral shift amount, This is the maximum spectral shift.

9. The long-distance OFDR fast demodulation method according to claim 1, characterized in that, In step 4, the steps for constructing the corresponding inference sample data based on the local reference real signal and the local measurement real signal of each sensing area are as follows: Step 41: Perform mean removal and standardization processing on the local reference real signals and local measurement real signals of each sensing area to obtain multiple sets of standardized reference real signals and standardized measurement real signals with the same amplitude scale. The standardized real reference signal for the i-th sensing region is shown below: The standardized measurement data of the i-th sensing region is shown below: in, For the local reference real signal of the i-th sensing region The mean, For the local reference real signal of the i-th sensing region Standard deviation; For the local measurement real signal of the i-th sensing region The mean, For the local measurement real signal of the i-th sensing region The standard deviation; ε is a local constant used to avoid the denominator being zero; Step 42: Stack the standardized reference real signal and the standardized measurement real signal of each sensing region in a dual-channel configuration to obtain a two-dimensional inference matrix for each sensing region. The two-dimensional inference matrix of the i-th sensing region is shown below: in, For the standardized reference real signal of the i-th sensing region, This represents the standardized measurement real signal for the i-th sensing region.

10. A long-range OFDR fast demodulation device based on synthetic data-trained neural networks, comprising a processor and a memory connected together, wherein the memory stores a computer program and a neural network model for execution by the processor, characterized in that, When the processor executes the computer program, it invokes the neural network model to perform the long-distance OFDR fast demodulation method as described in claim 1.

Images