Convolution kernel prediction method for optical fiber signals based on adaptive Fourier deconvolution model

Abstract

Aiming at the processing of fiber Bragg grating sensor signals, the present invention discloses a fiber signal convolution kernel prediction method based on an adaptive Fourier deconvolution model, including: obtaining reflection spectrum data of different central wavelengths; setting convolution kernels with different parameters; convolving the two to obtain simulated spectrum data, converting the time domain and frequency domain signals of the simulated spectrum data into two-dimensional images through GAF coding technology and splicing them to expand the correlation between data; constructing an adaptive Fourier deconvolution neural network model; the deconvolution neural network takes the spliced ​​image as input and the parameters of the convolution kernel as output; after the training is completed, the fiber Bragg grating sensor reflection spectrum signal is collected by a demodulator, and the system response of the demodulator is determined by the model. The convolution kernel obtained by the present invention can be used for the restoration of real fiber spectrum signals, which will greatly improve the accuracy of subsequent various feature extractions of the fiber Bragg grating demodulator, thereby improving its applicability and stability in complex signal environments.

Description

Technical Field

[0001] The invention belongs to the technical field of optical fiber sensing, and in particular relates to an optical fiber signal convolution kernel prediction method based on an adaptive Fourier deconvolution model. Background Art

[0002] Acquiring the real spectral signal of the fiber Bragg grating sensor is the core of the fiber Bragg grating sensing system, which is directly related to the measurement accuracy of the fiber Bragg grating sensing system. With the rapid development of fiber wireless communication technology and fiber sensor networks, the demand for high-precision reconstruction and demodulation of complex fiber spectral signals is increasing. Due to the convolution effect of the demodulator system on the spectral signal of the fiber Bragg grating sensor, the demodulator system is often unable to obtain the real spectral signal of the fiber Bragg grating sensor, which limits the accuracy of the subsequent feature extraction (such as the peak center) of the demodulation system. In addition to the convolution effect of the demodulator system, the spectral signal of the fiber Bragg grating sensor actually obtained may contain multiple frequency components, noise interference, and time-varying characteristics, which makes it difficult for general demodulation technology to obtain stable and accurate demodulation results.

[0003] In order to obtain the real spectral data of the fiber sensor, the traditional demodulation method usually uses a simulated convolution kernel to deconvolve the reflected spectrum signal. However, this method often shows limitations when facing complex fiber Bragg grating sensor signals, such as the inability to restore the real waveform, poor anti-interference ability, and waveform distortion. Therefore, if the system response of the demodulation system under different states can be accurately obtained, that is, the convolution kernel signal of the reflected spectrum, it is very important to restore the real spectral signal and improve the system performance and robustness.

[0004] Deconvolution algorithms based on Fourier transform are often used to recover convolution signals and extract convolution kernel signals. As a powerful frequency domain analysis tool, Fourier transform can convert signals from time domain to frequency domain and reveal the hidden spectral features in the signal; deconvolution technology can restore signals distorted by system response, making signal reconstruction more accurate. However, traditional deconvolution technology based on Fourier transform has problems with parameter fixity and insufficient adaptability when dealing with real-time changing fiber Bragg grating sensor signals. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a method for predicting the convolution kernel of optical fiber signals based on an adaptive Fourier deconvolution model. The method provides an adaptive Fourier deconvolution model, and introduces GAF coding technology to utilize the correlation between data. In addition, the model is trained and optimized by using deep learning technology and combining Fourier deconvolution (FourierDeconvolution) and adaptive algorithms through simulated data sets and real-time collected data sets, which solves the problems that traditional demodulation technology may face when processing complex signals, such as unknown system response and insufficient analysis accuracy caused by noise interference. The model can not only dynamically adjust the parameters and structures in the model according to different signal characteristics, but also cope with changes in signal strength at different frequencies, and has strong adaptability, which improves the ability of the demodulation system to restore real spectral signals in complex signal environments.

[0006] The technical solution adopted by the present invention is:

[0007] According to a first aspect of the present invention, a method for predicting a convolution kernel of an optical fiber signal based on an adaptive Fourier deconvolution model is provided, the method comprising:

[0008] Acquire reflection spectrum data of different central wavelengths; the acquired reflection spectrum data is the spectrum signal of the fiber Bragg grating sensor, rather than the system response signal of the demodulator;

[0009] Use the Gaussian kernel as the convolution kernel and set the convolution kernel with different parameters;

[0010] The convolution kernel is used as the system response of the demodulator, and the convolution kernels with different parameters are convolved with the reflection spectrum data of different central wavelengths to obtain the simulated spectrum data of the demodulator, which is used to simulate the system response signal of the demodulator;

[0011] Data preprocessing: converting the time domain and frequency domain signals of the simulated spectral data into two-dimensional images through GAF coding technology, and splicing the two-dimensional images to obtain a spliced ​​image;

[0012] An adaptive Fourier deconvolution model is constructed, which includes a deconvolution neural network. The deconvolution neural network takes the two-dimensional image data of the spliced ​​time domain and frequency domain signals as input, extracts useful features from the image, and takes the parameters of the convolution kernel as output.

[0013] Using the simulated spectrum data of the demodulator and the corresponding convolution kernel, the adaptive Fourier deconvolution model is trained to obtain an optimal deconvolution neural network;

[0014] Finally, the reflected spectrum signal of the fiber Bragg grating sensor is collected by a demodulator and converted into a spliced ​​image. The corresponding convolution kernel parameters are obtained by using an adaptive Fourier deconvolution model to determine the system response of the demodulator.

[0015] Furthermore, the reflectance spectrum data of different central wavelengths are obtained, including:

[0016] According to the principle of FBG sensor, the mathematical model of fiber Bragg grating reflection spectrum is obtained;

[0017] Based on the mathematical model of fiber Bragg grating reflection spectrum, the reflection spectrum data of different central wavelengths are obtained.

[0018] Furthermore, the mathematical model of fiber Bragg grating reflection spectrum includes:

[0019] The mathematical model of fiber Bragg grating reflection spectrum is the approximate Gaussian distribution function of fiber Bragg grating reflection spectrum, and its wavelength value is Corresponding light intensity The expression is:

[0020]

[0021] Where: is the reflected spectrum intensity amplitude, is the central wavelength of the reflected light, is the 3dB bandwidth of the reflection spectrum.

[0022] Furthermore, the Gaussian kernel function It is expressed as:

[0023]

[0024] in, is the input vector, π is the circumference of a circle, is the dimension of the vector, is the mean vector, is the covariance matrix; , and are the parameters of the convolution kernel.

[0025] Further, GAF coding includes:

[0026]

[0027] In the formula, and are two orthogonal sequences; is a set of time series data with a length of N, namely, the input spectral data;

[0028] Calculate the Gram matrix G, defined as:

[0029]

[0030] Transform the elements of matrix G through the hyperbolic tangent function to obtain the GAF image A:

[0031]

[0032] In the formula, is the pixel value of the GAF image, are the elements of the Gram matrix, i and j It's the location.

[0033] Furthermore, GAF is normalized before encoding:

[0034]

[0035] In the formula, is the input spectral data, yes The data in is the normalized data.

[0036] Furthermore, the deconvolutional neural network is a neural network model including a deconvolution layer, an activation function and a regularization layer.

[0037] Further, the training includes:

[0038] The spliced ​​image corresponding to the simulated spectral data is used as the input of the adaptive Fourier deconvolution model to output the corresponding convolution kernel parameters;

[0039] Determine the convolution kernel used when acquiring the simulated spectral data, and calculate the error between the two;

[0040] If the error is within the set range, the training is completed, otherwise the deconvolution neural network in the adaptive Fourier deconvolution model continues to be optimized.

[0041] Furthermore, the method also includes:

[0042] Verify the adaptive Fourier deconvolution model, including:

[0043] The demodulator is directly connected to the spectrometer, and the spectral waveform curve measured by the spectrometer represents the spectral information inside the demodulator, that is, the system response of the demodulator, so as to obtain the corresponding convolution kernel parameters; at the same time, the reflection spectrum signal collected by the demodulator is converted into a spliced ​​image, and the corresponding convolution kernel parameters are obtained using the adaptive Fourier deconvolution model;

[0044] The parameter errors between the two convolution kernels are compared to verify the adaptive Fourier deconvolution model.

[0045] According to a second aspect of the present invention, a demodulator is provided, which adopts any one of the above-mentioned methods for predicting convolution kernels of optical fiber signals based on an adaptive Fourier deconvolution model.

[0046] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0047] The present invention utilizes the correlation between data to add GAF ​​coding, converts the one-dimensional spectral data under different stresses collected by the demodulator into two-dimensional data, and uses the obtained two-dimensional data as the input of the neural network, which can enhance the degree of fusion between different spectral data, making the convolution kernel obtained by the neural network more accurate. After obtaining the deconvolution kernel (i.e., the response function of the system), the adaptive Fourier deconvolution model is used to minimize the distortion effect and effectively restore the accurate information of the original signal. This method is not only applicable to the field of optical fiber signal processing, but also can be applied in the fields of image processing, audio processing, etc., providing powerful tools and technical support for improving the accuracy and quality of data recovery and processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is the flow chart of the adaptive Fourier deconvolution model based on GAF coding;

[0049] Figure 2 This is the flow chart of Fourier deconvolution algorithm;

[0050] Figure 3 To use GAF coding to convert the time domain and frequency domain signals of different strain waveforms into a two-dimensional image schematic diagram;

[0051] Figure 4 Comparison chart between adaptive Fourier deconvolution model and traditional deconvolution model. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0053] The present invention provides a method for predicting the convolution kernel of an optical fiber signal based on an adaptive Fourier deconvolution model, and relates to an adaptive Fourier deconvolution model and its application in a demodulator. The adaptive Fourier deconvolution model introduces GAF coding technology, uses deep learning technology, combines Fourier deconvolution and adaptive algorithms (the convolution kernels of different data are different, and the deconvolution model can be adjusted in real time to output the convolution kernels of different signals), trains and optimizes the model through simulated data sets and real-time collected signals, and solves the problems that traditional demodulation technology may face when processing complex signals, such as unknown system response and insufficient analysis accuracy caused by noise interference. The model can not only dynamically adjust the parameters and structures in the model according to different signal characteristics, but also cope with changes in signal strength of different frequencies, and has strong adaptability, thereby improving the applicability and stability of the demodulator in a complex signal environment.

[0054] In a fiber Bragg grating sensor, incident light is reflected by the grating in the optical fiber to generate reflected light. However, in the demodulator system, the optical signal modulated by the FBG sensor received by the demodulator is not the real FBG signal, but the result of the convolution of the system scanning laser spectrum and the characteristic spectrum of the FBG sensor. Based on this, the present invention first obtains the fiber Bragg grating sensor reflection spectrum signal by two methods: simulation data generation and demodulator acquisition. Then, GAF coding is used to convert the one-dimensional data into two-dimensional image data. Secondly, an adaptive Fourier deconvolution model based on deep learning is proposed, and it is trained and optimized. Finally, the GAF image data is imported into the trained model to obtain the convolution kernel parameters corresponding to the reflection spectrum, so as to determine the response of the demodulator system.

[0055] Among them, building and training an adaptive Fourier deconvolution model based on deep learning includes the following steps:

[0056] Based on the principle of fiber Bragg grating sensor, the mathematical model of reflected spectrum signal is calculated and the simulation data set is constructed;

[0057] The simulated data set data is preprocessed, specifically: the time domain and frequency domain signals of the simulated spectral data are converted into two-dimensional images by GAF coding technology, and the two-dimensional images are spliced ​​to obtain a spliced ​​image;

[0058] A deconvolutional neural network is built based on a deep learning framework. The deconvolutional neural network takes the spliced ​​image as input and the parameters of the convolution kernel as output.

[0059] The simulated data set is divided into a training set and a test set. The deconvolution neural network is continuously optimized through the training set to output the convolution kernel of the convolution signal.

[0060] Compare the output convolution kernel with the original convolution kernel parameters until the error is within the allowable range.

[0061] Specifically, the optical fiber signal convolution kernel prediction method based on the adaptive Fourier deconvolution model of the present invention is as follows: Figure 1 As shown, the following steps are included:

[0062] Step 1: According to the principle of FBG sensor, calculate its reflection spectrum mathematical model.

[0063] Fiber Bragg Grating (FBG) sensor is a sensor that uses fiber optic technology to measure physical quantities such as temperature, stress and pressure. Its principle is based on the Bragg grating in the optical fiber, which is a structure that periodically changes the refractive index in the core of the optical fiber. When light passes through the grating, it will produce a strong reflection at a specific wavelength, which is called the Bragg wavelength ( ), given by the period of the grating ( ) and the effective refractive index of the optical fiber ( ) decide, follow the formula .

[0064] When external factors such as temperature or stress act on the optical fiber, the period or refractive index of the grating will change, causing the Bragg wavelength to shift. By measuring the wavelength change of the reflected light with a spectrum analyzer, the change in physical quantity can be accurately determined.

[0065] The Bragg wavelength of the FBG sensor ( ) With the strain ( ) changes, and the relationship can be described by the following formula:

[0066]

[0067] in, Is in strain The Bragg wavelength is is the initial strain The Bragg wavelength. is the strain sensitivity coefficient, which expresses the rate of change of the Bragg wavelength per unit strain. is the current strain, is the initial strain.

[0068] For example, assuming the initial strain is 0, the initial Bragg wavelength is 1550nm, strain sensitivity coefficient for For strain , the Bragg wavelength is:

[0069]

[0070] For strain , the Bragg wavelength is:

[0071]

[0072] In order to train the model of the present invention, it is necessary to generate reflection spectrum data of different central wavelengths, and then to obtain the required simulated spectrum data through convolution.

[0073] The mathematical model of the fiber Bragg grating reflection spectrum is the approximate Gaussian distribution function of the fiber Bragg grating reflection spectrum. Corresponding light intensity The expression is:

[0074]

[0075] Where: is the reflected spectrum intensity amplitude, is the central wavelength of the reflected light, is the 3dB bandwidth of the reflection spectrum.

[0076] The required simulation data can be generated according to the above formula.

[0077] Step 2: Define the convolution kernel function.

[0078] The fiber Bragg grating reflection spectrum signal collected in real time can be approximately described as a Gaussian signal, and the characteristic function of the demodulator system is similar to the Gaussian distribution. Therefore, the present invention uses the Gaussian kernel as the convolution kernel of the simulated data set, and simulates the data collected by the demodulator by convolving the simulated reflection spectrum data of different central wavelengths with the Gaussian kernel. Gaussian kernel function It can be expressed as:

[0079]

[0080] in, is the input vector, is the dimension of the vector, is the mean vector, is the covariance matrix.

[0081] Step 3: Construct a simulation data set based on the mathematical model of fiber Bragg grating reflection spectrum and Gaussian kernel function.

[0082] In this embodiment, the central wavelength is set In the 1542-1548nm range, the wavelength resolution is 10pm, so there are 600 points in the entire range. The spectral amplitude range is 0.6-1.0mw, is 0.2-0.35nm. The convolution kernel is , , The three parameters affect each parameter, and each parameter varies from 0 to 100 with an interval of 5. There are 20×20×20=8000 sets of convolution kernels with different parameters. The convolution kernel is convolved with the reflectance spectrum data of different central wavelengths to obtain 4,800,000 sample data, which is the simulated data set of the model.

[0083] Step 4: Get the real data set collected by the demodulator.

[0084] In this embodiment, a demodulator is used to obtain the spectral data of the sensor under different central wavelengths and different noise conditions. The central wavelength range is 1542-1548nm. At each wavelength, no less than 10 spectra were collected, and the spectral characteristics under different noises were reflected as much as possible.

[0085] Real data sets are also used in the training, mainly for validation and testing of the model. For example, the demodulator system response can be directly measured using a spectrometer and compared with the system response obtained by the deconvolution model.

[0086] Step 5: Data preprocessing.

[0087] The simulated spectral data and the real data set are converted and normalized. The format conversion uses the GAF algorithm, such as Figure 3 As shown, the specific formula is as follows:

[0088]

[0089] in, is a set of time series data with a length of N, namely, the input simulated spectral data; and are two orthogonal sequences.

[0090] Compute the Gram matrix, defined as:

[0091]

[0092] Transform the elements of matrix G through the hyperbolic tangent function to obtain the GAF image A:

[0093]

[0094] in, is the pixel value of the GAF image, are the elements of the Gram matrix.

[0095] Before format conversion, the data can be normalized to limit the light intensity to the interval [0,1]. The formula used is:

[0096]

[0097] In this embodiment, both the time domain and frequency domain signals of the simulated spectral data are converted into two-dimensional images through the GAF coding technology, and the time domain and frequency domain images are spliced ​​as the input of the Fourier deconvolution model.

[0098] Step 6: Build a deconvolution neural network to extract the convolution kernel of the convolution signal.

[0099] 1. Determine the structure of the network. Choose the appropriate network depth and width based on the characteristics of the task and data.

[0100] 2. Define the model architecture. Import the PyTorch deep learning framework and create a neural network model that includes deconvolution layers, activation functions, and regularization layers (such as batch normalization).

[0101] 3. Add a deconvolution layer. In PyTorch, you can use "torch.nn.ConvTranspose2d".

[0102] 4. Add an activation function layer. Usually ReLU is used to introduce nonlinear characteristics and increase the expressive power of the network.

[0103] 5. Define the loss function and optimizer. The loss function is used to measure the specific metric of the quality of convolution kernel extraction, and the optimizer selects the Adam optimization algorithm.

[0104] 6. Compile the model. Define the loss function and optimizer evaluation metrics.

[0105] 7. Train the model. Use the training set to train the model.

[0106] 8. Model evaluation and prediction. Use the test set to evaluate the performance of the model.

[0107] 9. Optimize parameters and improve structure.

[0108] It should be noted that the deconvolution neural network used to extract the convolution kernel of the convolution signal is not limited to the neural network model in this embodiment.

[0109] Step 7: Use Fourier deconvolution algorithm to further restore the convolution kernel signal.

[0110] The Fourier deconvolution algorithm converts the convolution operation into a multiplication operation by using Fourier transform to restore the original signal. In signal processing, convolution refers to convolution of a signal with a convolution kernel (or system function) to obtain a new signal. This process can be expressed by the convolution theorem:

[0111]

[0112] in, is the observed convolution signal, is the original signal, is the convolution kernel, Represents the convolution operation. Performing Fourier transform on the signal and convolution kernel respectively, we can get:

[0113]

[0114] in, , , They are , , Fourier transform. In order to restore the original signal , perform the inverse operation of the multiplication operation in the Fourier transform domain, that is, perform deconvolution. The specific formula is as follows:

[0115]

[0116] in, is the Fourier transform of the convolution kernel. Perform inverse Fourier transform to obtain the restored original signal .

[0117] The present invention uses the convolution kernel output by the deconvolution neural network as the system response of the entire demodulator system. When the error is less than the allowable range, the model training is completed. If the error exceeds the allowable range, it is necessary to continue to optimize the deconvolution neural network. Figure 2 shown.

[0118] Step 8: Verify the feasibility of the adaptive Fourier deconvolution model through real-time fiber Bragg grating sensor reflection spectrum data.

[0119] 1. Build a fiber Bragg grating demodulator experimental platform, including fiber Bragg grating sensor, equal strength cantilever beam, weight, power supply, and PC terminal.

[0120] 2. Develop PC-side fiber Bragg grating sensor data acquisition software and incorporate the etalon grating data. Compare the acquired data with the etalon grating data and output the center wavelength value of the sensor data in real time. Adjust the internal galvanometer structure of the demodulator according to the center wavelength value, output the corresponding spectral information, and use the spectrometer to measure its spectral waveform curve. The spectral waveform curve is also a peak similar to a Gaussian kernel, so the corresponding convolution kernel parameters can be obtained.

[0121] 3. Export the fiber Bragg grating sensor data collected by the software and input it into the adaptive Fourier deconvolution model to obtain its corresponding convolution kernel signal

[0122] 4. Compare the results processed by the adaptive Fourier deconvolution model with the results measured by the spectrometer to verify the feasibility of the model.

[0123] Figure 4 This is a comparison chart between the adaptive Fourier deconvolution model and the traditional deconvolution model. GAF coding is used to convert one-dimensional spectral data into two-dimensional data, and it can achieve adaptability.

[0124] The present invention also provides a demodulator, which adopts any one of the above-mentioned methods for predicting the convolution kernel of optical fiber signals based on the adaptive Fourier deconvolution model.

[0125] In summary, the present invention utilizes GAF coding to convert the one-dimensional spectral data under different stresses collected by the demodulator into two-dimensional data. The obtained two-dimensional data is used as the input of the neural network, which can enhance the degree of fusion between different spectral data, so that the convolution kernel obtained by the neural network is more accurate. After obtaining the deconvolution kernel (i.e., the response function of the system), the accurate information of the original signal can be effectively restored. This method is not only applicable to the field of signal processing, but also can be applied in the fields of image processing, audio processing, etc., and provides powerful tools and technical support for improving the accuracy and quality of data recovery and processing.

[0126] It should be noted that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0127] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, and two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0128] It will be easily understood by those skilled in the art that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting convolution kernel of optical fiber signal based on adaptive Fourier deconvolution model, characterized in that: The method includes: Obtain reflection spectrum data at different central wavelengths; Use the Gaussian kernel as the convolution kernel and set the convolution kernel with different parameters; The convolution kernel is used as the system response of the demodulator, and the convolution kernels with different parameters are convolved with the reflection spectrum data of different central wavelengths to obtain the simulated spectrum data of the demodulator; The time domain and frequency domain signals of the simulated spectral data are converted into two-dimensional images by GAF coding technology, and the two-dimensional images are spliced ​​to obtain a spliced ​​image; An adaptive Fourier deconvolution model is constructed, wherein the adaptive Fourier deconvolution model includes a deconvolution neural network; the deconvolution neural network takes the spliced ​​image as input and takes the parameters of the convolution kernel as output; Using the simulated spectrum data of the demodulator and the corresponding convolution kernel, the adaptive Fourier deconvolution model is trained to obtain an optimal deconvolution neural network; The reflected spectrum signal of the fiber Bragg grating sensor is collected by a demodulator and converted into a spliced ​​image. Then, the adaptive Fourier deconvolution model is used to obtain the corresponding convolution kernel parameters and determine the system response of the demodulator.

2. The method for predicting the convolution kernel of an optical fiber signal based on an adaptive Fourier deconvolution model according to claim 1, characterized in that: Acquire reflectance spectrum data at different central wavelengths, including: According to the principle of FBG sensor, the mathematical model of fiber Bragg grating reflection spectrum is obtained; Based on the mathematical model of fiber Bragg grating reflection spectrum, the reflection spectrum data of different central wavelengths are obtained.

3. The method for predicting the convolution kernel of an optical fiber signal based on an adaptive Fourier deconvolution model according to claim 2, characterized in that: Mathematical model of fiber Bragg grating reflection spectrum, including: The mathematical model of fiber Bragg grating reflection spectrum is the approximate Gaussian distribution function of fiber Bragg grating reflection spectrum, and its wavelength value is Corresponding light intensity The expression is: Where: is the reflected spectrum intensity amplitude, is the central wavelength of the reflected light, is the 3dB bandwidth of the reflection spectrum.

4. The method for predicting the convolution kernel of an optical fiber signal based on an adaptive Fourier deconvolution model according to claim 1, characterized in that: Gaussian kernel function It is expressed as: in, is the input vector, π is the circumference of a circle, is the dimension of the vector, is the mean vector, is the covariance matrix; , and are the parameters of the convolution kernel.

5. The method for predicting the convolution kernel of an optical fiber signal based on an adaptive Fourier deconvolution model according to claim 1, characterized in that: GAF codes include: In the formula, and are two orthogonal sequences; is a set of time series data with a length of N, namely, the input spectral data; Calculate the Gram matrix G, defined as: Transform the elements of matrix G through the hyperbolic tangent function to obtain the GAF image A: In the formula, is the pixel value of the GAF image, are the elements of the Gram matrix, i and j It's the location.

6. The method for predicting convolution kernel of optical fiber signal based on adaptive Fourier deconvolution model according to claim 1, characterized in that: GAF is normalized before encoding: In the formula, is the input spectral data, yes The data in is the normalized data.

7. The method for predicting convolution kernel of optical fiber signal based on adaptive Fourier deconvolution model according to claim 1, characterized in that: The deconvolutional neural network is a neural network model that includes a deconvolution layer, an activation function, and a regularization layer.

8. The method for predicting convolution kernel of optical fiber signal based on adaptive Fourier deconvolution model according to claim 1, characterized in that: Training includes: The spliced ​​image corresponding to the simulated spectral data is used as the input of the adaptive Fourier deconvolution model to output the corresponding convolution kernel parameters; Determine the convolution kernel used when acquiring the simulated spectral data, and calculate the error between the convolution kernel and the convolution kernel output by the adaptive Fourier deconvolution model; If the error is within the set range, the training is completed, otherwise the deconvolution neural network in the adaptive Fourier deconvolution model continues to be optimized.

9. The method for predicting convolution kernel of optical fiber signal based on adaptive Fourier deconvolution model according to claim 1, characterized in that: The method further includes: Verify the adaptive Fourier deconvolution model, including: The demodulator is directly connected to the spectrometer. The spectral waveform curve measured by the spectrometer represents the spectral information inside the demodulator, that is, the system response of the demodulator, so as to obtain the corresponding convolution kernel parameters; At the same time, the reflection spectrum signal collected by the demodulator is converted into a spliced ​​image, and the corresponding convolution kernel parameters are obtained using the adaptive Fourier deconvolution model; The parameter errors between the two convolution kernels are compared to verify the adaptive Fourier deconvolution model.

10. A demodulator, characterized in that: The demodulator adopts the optical fiber signal convolution kernel prediction method based on the adaptive Fourier deconvolution model as described in any one of claims 1 to 9.

Images