A method, system and computer readable medium for Brillouin gain spectrum reduction
Patent Information
- Application Number
- CN202410110462.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-26
- Publication Date
- 2026-09-25
- Estimated Expiration
- 2044-01-26
AI Technical Summary
在大多数情况下，在适当的频率范围内重新测量布里渊增益谱是不切实际的
[0043]经由上述的技术方案可知，与现有技术相比，本发明公开提供了一种布里渊增益谱还原方法、系统及可存储介质，通过对带噪残缺布里渊频谱进行还原降噪，即使当分布式传感系统散射光信号的频谱图中低完整度和高信噪比时，也能快速准确地获得传感信息，还原监督变分自编码器网络对完整度具有更好的容忍度，可以使得整个传感系统的测量速度和测量精度的得到明显提升。
Smart Images
Figure CN117928614B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fiber optic sensing technology, and more specifically to a Brillouin gain spectrum restoration method, system, and storage medium. Background Technology
[0002] Currently, distributed Brillouin fiber optic sensing technology has significant application value in fields such as structural health monitoring of large-scale infrastructure and geological and geophysical research. In particular, the Brillouin optical time-domain analyzer based on stimulated Brillouin scattering can achieve real-time and high-speed distributed measurement of strain and temperature over hundreds of kilometers of fiber with only one optical fiber. It has advantages such as no continuous blind zone, a large number of sensing points, and long measurement distance, which can significantly reduce the complexity of the entire monitoring system and also reduce the construction cost of the monitoring system.
[0003] However, in long-term monitoring of actual engineering projects, traditional distributed Brillouin fiber optic sensing technology suffers from low spatial resolution due to limitations in its sensing mechanism, severely restricting its application areas and practical value. Due to structural strain and temperature variations, the set scanning frequency range often fails to cover the peak information of the Brillouin Gain Spectrum (BGS), resulting in incomplete BGS measurements and incomplete Brillouin gain spectra. Consequently, important information such as Brillouin Frequency Shift (BFS) and peak power cannot be extracted. Furthermore, while techniques such as ramp-assisted sensing, optical linear frequency modulation chains, and frequency combs can improve sensing speed, for real-time signal acquisition scenarios with high response speed requirements, excessively low sampling frequencies may lead to the loss of high-frequency signals, making traditional methods unable to fit the data. In addition, when external interference is short-lived, increasing the frequency scanning interval and reducing the average number of signals will prevent the sensing system from sensing the target's state in a timely manner, resulting in incomplete Brillouin gain spectra. Some operational errors in actual engineering projects may also lead to incomplete Brillouin gain spectra. In most cases, remeasuring the Brillouin gain spectrum within the appropriate frequency range is impractical. Therefore, reconstructing the complete Brillouin gain spectrum from the incomplete Brillouin gain spectrum and extracting the Brillouin frequency shift are key technologies for Brillouin distributed fiber optic sensors. Currently, there are no existing technologies using neural networks to recover the incomplete Brillouin gain spectrum and extract information.
[0004] Therefore, how to provide a Brillouin gain spectrum restoration method that can solve the above problems is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] In view of this, the present invention provides a Brillouin gain spectrum restoration method, system and storage medium. By constructing a restoration supervised variational autoencoder network, the noisy and incomplete Brillouin gain spectrum is restored and denoised. Based on the restored Brillouin gain spectrum, sensing information can be obtained quickly and accurately, so as to significantly improve the measurement speed and measurement accuracy of the entire sensing system.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A Brillouin gain spectrum restoration method includes the following steps:
[0008] Obtain complete noise-free Brillouin gain spectrum data and incomplete noise-free Brillouin gain spectrum data, and process the complete noise-free Brillouin gain spectrum data and the incomplete noise-free Brillouin gain spectrum data to obtain complete noise-brillouin gain spectrum data and incomplete noise-brillouin gain spectrum data.
[0009] Construct a variational autoencoder network and a supervised variational autoencoder network, wherein the variational autoencoder network includes a first encoder and a first decoder, and the supervised variational autoencoder network includes a second encoder and a second decoder;
[0010] The variational autoencoder network is optimized using complete noisy Brillouin gain spectrum data, and the supervised variational autoencoder network is optimized using incomplete noisy Brillouin gain spectrum data.
[0011] The optimized second encoder and the first decoder are selected to form a reductive supervised variational autoencoder network;
[0012] The noisy, incomplete Brillouin gain spectrum is processed using the aforementioned restored supervised variational autoencoder network to obtain complete Brillouin gain spectrum data.
[0013] Preferably, the process of optimizing the variational autoencoder network using complete noisy Brillouin gain spectrum data includes:
[0014] The complete noisy Brillouin gain spectrum data is input into the first encoder for encoding processing to obtain the corresponding first type of Gaussian distribution data, as well as the mean and standard deviation of the first type of Gaussian distribution data;
[0015] The first type of Gaussian distribution data is randomly sampled to obtain the first random data, and the first random data is input into the first decoder for decoding processing to obtain the corresponding first Brillouin gain spectrum data;
[0016] The first total loss value is obtained based on the first Brillouin gain spectrum data, the complete noise-free Brillouin gain spectrum data, and the first type of Gaussian distribution data.
[0017] The variational autoencoder network is optimized based on the first total loss value. Optimization stops when the number of optimization attempts meets the first preset number of attempts or when the first Brillouin gain spectrum data meets the preset quality requirements.
[0018] Preferably, the specific process for optimizing the supervised variational autoencoder network using incomplete noisy Brillouin gain spectrum data includes:
[0019] The incomplete noisy Brillouin gain spectrum data is input into the second encoder for encoding processing to obtain the corresponding second type of Gaussian distribution data, as well as the mean and standard deviation of the second type of Gaussian distribution data;
[0020] Random sampling is performed on the second type of Gaussian distribution data to obtain second random data, and the second random data is input into the second decoder for decoding processing to obtain the corresponding second Brillouin gain spectrum data;
[0021] The second total loss value is obtained based on the second Brillouin gain spectrum data, the complete noise-free Brillouin gain spectrum data, and the second type of Gaussian distribution data.
[0022] The variational autoencoder network is optimized based on the second total loss value. Optimization stops when the number of optimization attempts meets the second preset number of attempts or when the second Brillouin gain spectrum data meets the preset quality requirements.
[0023] Preferably, the specific processing procedure for processing the noisy, incomplete Brillouin gain spectrum using the reduced-supervised variational autoencoder network to obtain complete Brillouin gain spectrum data includes:
[0024] The noisy, incomplete Brillouin gain spectrum is input into the second encoder for encoding processing to obtain the corresponding latent features;
[0025] The latent features are input into the first decoder to obtain a noise-free complete gain spectrum.
[0026] Preferably, the specific processing steps for obtaining the first total loss value include:
[0027] Substitute the first Brillouin gain spectrum data and the corresponding complete noise-free Brillouin gain spectrum data into the loss function to obtain the first reconstruction loss;
[0028] The KL divergence of the first type of Gaussian distribution data and the standard Gaussian distribution is used to obtain the corresponding first distribution difference loss.
[0029] The first total loss value is obtained by summing the first distribution difference loss and the first reconstruction loss.
[0030] Preferably, the specific processing steps for obtaining the second total loss value include:
[0031] Substituting the second Brillouin gain spectrum data and the corresponding complete noise-free Brillouin gain spectrum data into the loss function, we obtain the second reconstruction loss;
[0032] Perform KL divergence analysis between the second type of Gaussian distribution data and the first type of Gaussian distribution data to obtain the corresponding second distribution difference loss;
[0033] The second total loss value is obtained by summing the second distribution difference loss and the second reconstruction loss.
[0034] Preferably, the second decoder includes a transposed convolutional layer, an attention module, and a fully connected layer, and the first encoder includes a convolutional layer, an attention module, and a fully connected layer.
[0035] Preferably, the reduction-supervised variational autoencoder network is evaluated using the coefficient of determination, root mean square error, and uncertainty.
[0036] The present invention also provides a reconstruction system utilizing the Brillouin gain spectrum restoration method described in any one of the preceding claims, comprising:
[0037] The data processing module is used to acquire complete noise-free Brillouin gain spectrum data and incomplete noise-free Brillouin gain spectrum data, and process the complete noise-free Brillouin gain spectrum data and the incomplete noise-free Brillouin gain spectrum data to obtain complete noise-brillouin gain spectrum data and incomplete noise-brillouin gain spectrum data.
[0038] The first network construction module is used to construct a variational autoencoder network and a supervised variational autoencoder network, wherein the variational autoencoder network includes a first encoder and a first decoder, and the supervised variational autoencoder network includes a second encoder and a second decoder.
[0039] The network optimization module is used to optimize the variational autoencoder network using complete noisy Brillouin gain spectrum data, and simultaneously optimize the supervised variational autoencoder network using incomplete noisy Brillouin gain spectrum data.
[0040] The second network construction module is used to select the optimized second encoder and the first decoder to form a reductive supervised variational autoencoder network.
[0041] The restoration module is used to process the noisy, incomplete Brillouin gain spectrum using the restoration-supervised variational autoencoder network to obtain complete Brillouin gain spectrum data.
[0042] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the Brillouin gain spectrum restoration method as described in any of the preceding claims.
[0043] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a Brillouin gain spectrum restoration method, system and storage medium. By restoring and denoising the noisy and incomplete Brillouin spectrum, even when the spectrum of the scattered light signal of the distributed sensing system has low integrity and high signal-to-noise ratio, the sensing information can be obtained quickly and accurately. The restoration supervised variational autoencoder network has better tolerance for integrity, which can significantly improve the measurement speed and measurement accuracy of the entire sensing system. Attached Figure Description
[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0045] Figure 1 A flowchart of a Brillouin gain spectrum restoration method provided by the present invention;
[0046] Figure 2 The structural schematic diagram of the variational autoencoder network and the supervised variational autoencoder network provided by the present invention;
[0047] Figure 3 The structural principle diagram of the reduction-supervised variational autoencoder network provided by the present invention;
[0048] Figure 4 This is a schematic diagram of the structure of a residual attention convolutional neural network provided in an embodiment of the present invention.
[0049] Figure 5 The present invention provides a structural principle block diagram of a Brillouin gain spectrum restoration system. Detailed Implementation
[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0051] Example 1
[0052] See appendix Figure 1-3 As shown, this embodiment of the invention discloses a Brillouin gain spectrum restoration method, including the following steps:
[0053] Obtain complete noise-free Brillouin gain spectrum data and incomplete noise-free Brillouin gain spectrum data, and process the complete noise-free Brillouin gain spectrum data and incomplete noise-free Brillouin gain spectrum data to obtain complete noise-bearing Brillouin gain spectrum data and incomplete noise-bearing Brillouin gain spectrum data. This process can be achieved by adding Gaussian white noise to the complete noise-free Brillouin gain spectrum data and incomplete noise-bearing Brillouin gain spectrum data.
[0054] A variational autoencoder network (VAE) and a supervised variational autoencoder network (SVAE) are constructed. The variational autoencoder network (VAE) includes a first encoder and a first decoder, and the supervised variational autoencoder network (SVAE) includes a second encoder and a second decoder.
[0055] The variational autoencoder network (VAE) is optimized using complete noisy Brillouin gain spectrum data, while the supervised variational autoencoder network (SVAE) is optimized using incomplete noisy Brillouin gain spectrum data.
[0056] The optimized second encoder and the first decoder are selected to form the Restoration-Supervised Variational Autoencoder Network (RSVAE).
[0057] The noisy, incomplete Brillouin gain spectrum was processed using the Restoration-Supervised Variational Autoencoder (RSVAE) network to obtain complete Brillouin gain spectrum data.
[0058] In one specific embodiment, the process of optimizing the variational autoencoder network (VAE) using complete noisy Brillouin gain spectrum data includes:
[0059] The complete noisy Brillouin gain spectrum data is input into the first encoder for encoding processing to obtain the corresponding first type Gaussian distribution data, as well as the mean μ1 and standard deviation σ1 of the first type Gaussian distribution data;
[0060] The first type of Gaussian distribution data is randomly sampled to obtain the first random data, and the first random data is input into the first decoder for decoding processing to obtain the corresponding first Brillouin gain spectrum data;
[0061] The first total loss value is obtained based on the first Brillouin gain spectrum data, the complete noise-free Brillouin gain spectrum data, and the first type of Gaussian distribution data.
[0062] The variational autoencoder network is optimized based on the first total loss value. Optimization stops when the number of optimization attempts meets the first preset number of attempts or when the first Brillouin gain spectrum data meets the preset quality requirements.
[0063] Specifically, the optimization process for the variational autoencoder network based on the first total loss value includes:
[0064] The gradient of the first total loss value is calculated on the first Brillouin gain spectrum data, and the first Brillouin gain spectrum data is updated and optimized according to the backpropagation algorithm.
[0065] In a specific embodiment, the optimization process of the supervised variational autoencoder network SVAE using incomplete noisy Brillouin gain spectrum data includes:
[0066] The incomplete noisy Brillouin gain spectrum data is input into the second encoder for encoding processing to obtain the corresponding second type Gaussian distribution data, as well as the mean μ2 and standard deviation σ2 of the second type Gaussian distribution data.
[0067] Random sampling is performed on the second type of Gaussian distribution data to obtain second random data, and the second random data is input into the second decoder for decoding processing to obtain the corresponding second Brillouin gain spectrum data;
[0068] The second total loss value is obtained based on the second Brillouin gain spectrum data, the complete noise-free Brillouin gain spectrum data, and the second type of Gaussian distribution data.
[0069] The supervised variational autoencoder network SVAE is optimized based on the second total loss value. Optimization stops when the number of optimization attempts meets the second preset number of attempts or the second Brillouin gain spectrum data meets the preset quality requirements.
[0070] Specifically, the process of optimizing the supervised variational autoencoder network SVAE based on the second total loss value can be as follows:
[0071] The gradient of the second total loss value is calculated on the second Brillouin gain spectrum data, and the second Brillouin gain spectrum data is updated and optimized according to the backpropagation algorithm.
[0072] In a specific embodiment, the reduced supervised variational autoencoder RSVAE is used to process the noisy and incomplete Brillouin gain spectrum, and the specific processing process for obtaining complete Brillouin gain spectrum data includes:
[0073] Inputting the noisy and incomplete Brillouin gain spectrum into a second encoder for encoding processing to obtain corresponding latent features;
[0074] Inputting the latent features into a first decoder to obtain a noise-free complete gain spectrum.
[0075] Specifically, since the latent features learned from complete and incomplete curve profiles are different, the reconstructed missing values will be different from the true values. Label information of the unsupervised variational autoencoder is introduced and incorporated into the variational Bayesian inference (VBI) method to modify the network settings so that it can perform supervised learning. The label information is the posterior distribution in the learned complete BGS, which is used as the new prior distribution of the incomplete BGS for training. It can perform restoration for incomplete Brillouin gain spectra with high-frequency information loss, and can also perform noise reduction for complete Brillouin gain spectra.
[0076] In a specific embodiment, the specific processing process for obtaining the first total loss value includes:
[0077] Substituting the first Brillouin gain spectrum data and the corresponding complete noise-free Brillouin gain spectrum data into a loss function to obtain a first reconstruction loss;
[0078] Performing KL divergence on the first type of Gaussian distribution data and the standard Gaussian distribution to obtain a corresponding first distribution difference loss, wherein the specific expression for performing KL divergence is:
[0079]
[0080] In the formula, k is the dimension of z, represents complete noise-free Brillouin gain spectrum data, represents incomplete noise-free Brillouin gain spectrum data, μ(x (i) ) and σ(x (i) ) represent corresponding latent feature labels, k<n represents the latent feature corresponding to the complete noise-free Brillouin gain spectrum data, p represents the posterior distribution, and q represents the normal distribution;
[0081] Summing the first distribution difference loss and the first reconstruction loss to obtain the first total loss value.
[0082] In a specific embodiment, the specific processing process for obtaining the second total loss value includes:
[0083] Substituting the second Brillouin gain spectrum data and the corresponding complete noise-free Brillouin gain spectrum data into the loss function to obtain a second reconstruction loss;
[0084] Performing KL divergence on the second type of Gaussian distribution data and the first type of Gaussian distribution data to obtain the corresponding second distribution difference loss, and the specific expression is:
[0085]
[0086] In the formula, k is the dimension of z, represents complete noise-free Brillouin gain spectrum data, represents incomplete noise-free Brillouin gain spectrum data, μ(x (i) ) and σ(x (i) ) represent corresponding latent feature labels, k < n represents the latent feature corresponding to the complete noise-free Brillouin gain spectrum data, p represents the posterior distribution, and q represents the normal distribution;
[0087] Summing the second distribution difference loss and the second reconstruction loss to obtain a second total loss value.
[0088] In a specific embodiment, the second decoder comprises a transposed convolution layer, an attention module and a fully connected layer, and the first encoder comprises a convolution layer, an attention module and a fully connected layer.
[0089] Specifically, the first encoder and the second encoder have the same network structure, and the first decoder and the second decoder have the same network structure.
[0090] In a specific embodiment, the reduced supervised variational autoencoder network RSVAE is evaluated by using the coefficient of determination, root mean square error and uncertainty.
[0091] See the drawing Figure 5 , Embodiment 1 of the present invention also provides a reconstruction system using the Brillouin gain spectrum restoration method described in any one of the above embodiments, comprising:
[0092] a data processing module, configured to acquire complete noise-free Brillouin gain spectrum data and incomplete noise-free Brillouin gain spectrum data, and process the complete noise-free Brillouin gain spectrum data and the incomplete noise-free Brillouin gain spectrum data to obtain complete noisy Brillouin gain spectrum data and incomplete noisy Brillouin gain spectrum data;
[0093] The first network construction module is used to construct a variational autoencoder network (VAE) and a supervised variational autoencoder network (SVAE). The variational autoencoder network (VAE) includes a first encoder and a first decoder, and the supervised variational autoencoder network (SVAE) includes a second encoder and a second decoder.
[0094] The network optimization module is used to optimize the variational autoencoder network using complete noisy Brillouin gain spectrum data, and to optimize the supervised variational autoencoder network SVAE using incomplete noisy Brillouin gain spectrum data.
[0095] The second network construction module is used to select the optimized second encoder and the first decoder to form the Restoration Supervisory Variational Autoencoder Network (RSVAE).
[0096] The restoration module is used to process the noisy, incomplete Brillouin gain spectrum using the Restoration-Supervised Variational Autoencoder (RSVAE) network to obtain complete Brillouin gain spectrum data.
[0097] Embodiment 1 of the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the Brillouin gain spectrum restoration method as described in any of the above embodiments.
[0098] Example 2
[0099] The missing noisy Brillouin gain spectrum data is restored using the method provided in Example 1 to obtain complete noisy Brillouin gain spectrum data. Next, the complete noisy Brillouin gain spectrum data can be extracted to obtain the corresponding Brillouin frequency shift data. The specific process is as follows:
[0100] (1) Construct a residual attention convolutional neural network (RACNN). Input the complete, noise-free Brillouin gain spectrum data into the residual attention convolutional neural network for processing to obtain Brillouin frequency shift data. Calculate the frequency shift loss value based on the Brillouin frequency shift data and the real Brillouin frequency shift data. Update the Brillouin frequency shift according to the backpropagation algorithm. Continuously optimize the Brillouin frequency shift. Stop the optimization when the number of optimizations meets the set number of optimizations or the reconstructed Brillouin frequency shift meets the requirements.
[0101] (2) The Restoring and Extracting Convolutional Neural Network (RECNN) is composed of the Restoring-Supervised Variational Autoencoder Network (RSVAE) and the Residual Attention Convolutional Neural Network (RACNN). The noisy and incomplete Brillouin gain spectrum is fed into the RECNN to obtain the complete and noisy Brillouin gain spectrum and the Brillouin frequency shift optimized by the Residual Attention Convolutional Neural Network.
[0102] The Residual Attention Convolutional Neural Network (RACNN) includes: convolutional layers, max pooling layers, three attention modules, three residual modules (ResNet Blocks), a global average pooling layer, and a fully connected layer. See the appendix for the network structure. Figure 4 As shown, all convolutional layers and transposed convolutional layers are followed by Batch Normalization (BN) layers and SELU activation function layers;
[0103] The Restoration-Supervised Variational Encoder Network (RSVAE) is integrated with the Residual Attention Convolutional Neural Network (RACNN). By combining the attention mechanism and the residual module, the RSVAE is used to restore and denoise the Brillouin gain spectrum data with noisy arbitrary completeness, resulting in a noise-free complete Brillouin gain spectrum. Then, the Brillouin gain spectrum is processed by the Residual Attention Convolutional Neural Network to obtain the corresponding Brillouin frequency shift, which is then used for subsequent data processing.
[0104] Example 3
[0105] The reduced supervised variational encoder network (RSVAE) obtained in Example 1 was evaluated using the coefficient of determination, root mean square error, and uncertainty. The specific process of using the coefficient of determination for evaluation is as follows:
[0106] The coefficient of determination, R-squared, is a measure of the goodness of fit of the estimated regression equation. In linear regression, it is the ratio of the regression sum of squares to the total sum of squares, and its value is equal to the square of the correlation coefficient. For the R-squared index, simulation data was used to test its ability to follow changes in signal-to-noise ratio and completeness. The results are shown in Table 1.
[0107] Table 1 Results of the Coefficient of Determination
[0108]
[0109] The results show that RSVAE has better accuracy than other networks in terms of any degree of completeness, and its performance is better at lower degrees of completeness.
[0110] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0111] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for Brillouin gain spectrum reconstruction, characterized in that, Includes the following steps: Obtain complete noise-free Brillouin gain spectrum data and incomplete noise-free Brillouin gain spectrum data, and process the complete noise-free Brillouin gain spectrum data and the incomplete noise-free Brillouin gain spectrum data to obtain complete noise-brillouin gain spectrum data and incomplete noise-brillouin gain spectrum data. Construct a variational autoencoder network and a supervised variational autoencoder network, wherein the variational autoencoder network includes a first encoder and a first decoder, and the supervised variational autoencoder network includes a second encoder and a second decoder; The variational autoencoder network is optimized using complete noisy Brillouin gain spectrum data, and the supervised variational autoencoder network is optimized using incomplete noisy Brillouin gain spectrum data. The optimized second encoder and the first decoder are selected to form a reductive supervised variational autoencoder network; The noisy, incomplete Brillouin gain spectrum is processed using the aforementioned restored supervised variational autoencoder network to obtain complete Brillouin gain spectrum data.
2. The Brillouin gain spectrum reconstruction method according to claim 1, characterized in that, The process of optimizing the variational autoencoder network using complete noisy Brillouin gain spectrum data includes: The complete noisy Brillouin gain spectrum data is input into the first encoder for encoding processing to obtain the corresponding first type of Gaussian distribution data, as well as the mean and standard deviation of the first type of Gaussian distribution data; The first type of Gaussian distribution data is randomly sampled to obtain the first random data, and the first random data is input into the first decoder for decoding processing to obtain the corresponding first Brillouin gain spectrum data; The first total loss value is obtained based on the first Brillouin gain spectrum data, the complete noise-free Brillouin gain spectrum data, and the first type of Gaussian distribution data. The variational autoencoder network is optimized based on the first total loss value. Optimization stops when the number of optimization attempts meets the first preset number of attempts or when the first Brillouin gain spectrum data meets the preset quality requirements.
3. The Brillouin gain spectrum restoration method according to claim 2, characterized in that, The specific process of optimizing the supervised variational autoencoder network using incomplete noisy Brillouin gain spectrum data includes: The incomplete noisy Brillouin gain spectrum data is input into the second encoder for encoding processing to obtain the corresponding second type of Gaussian distribution data, as well as the mean and standard deviation of the second type of Gaussian distribution data; Random sampling is performed on the second type of Gaussian distribution data to obtain second random data, and the second random data is input into the second decoder for decoding processing to obtain the corresponding second Brillouin gain spectrum data; The second total loss value is obtained based on the second Brillouin gain spectrum data, the complete noise-free Brillouin gain spectrum data, and the second type of Gaussian distribution data. The supervised variational autoencoder network is optimized based on the second total loss value. Optimization stops when the number of optimization attempts meets the second preset number of attempts or when the second Brillouin gain spectrum data meets the preset quality requirements.
4. The Brillouin gain spectrum reconstruction method according to claim 1, characterized in that, The specific processing steps for processing the noisy, incomplete Brillouin gain spectrum using the aforementioned supervised variational autoencoder network to obtain complete Brillouin gain spectrum data include: The noisy, incomplete Brillouin gain spectrum is input into the second encoder for encoding processing to obtain the corresponding latent features; The latent features are input into the first decoder to obtain a noise-free complete gain spectrum.
5. The Brillouin gain spectrum restoration method according to claim 2, characterized in that, The specific process for obtaining the first total loss value includes: Substitute the first Brillouin gain spectrum data and the corresponding complete noise-free Brillouin gain spectrum data into the loss function to obtain the first reconstruction loss; The KL divergence of the first type of Gaussian distribution data and the standard Gaussian distribution is used to obtain the corresponding first distribution difference loss. The first total loss value is obtained by summing the first distribution difference loss and the first reconstruction loss.
6. The Brillouin gain spectrum restoration method according to claim 3, characterized in that, The specific process for obtaining the second total loss value includes: Substituting the second Brillouin gain spectrum data and the corresponding complete noise-free Brillouin gain spectrum data into the loss function, we obtain the second reconstruction loss; Perform KL divergence analysis between the second type of Gaussian distribution data and the first type of Gaussian distribution data to obtain the corresponding second distribution difference loss; The second total loss value is obtained by summing the second distribution difference loss and the second reconstruction loss.
7. The Brillouin gain spectrum restoration method according to claim 1, characterized in that, The second decoder includes a transposed convolutional layer, an attention module, and a fully connected layer, while the first encoder includes a convolutional layer, an attention module, and a fully connected layer.
8. The Brillouin gain spectrum restoration method according to claim 1, characterized in that, The reduction-supervised variational autoencoder network is evaluated using the coefficient of determination, root mean square error, and uncertainty.
9. A reconstruction system utilizing the Brillouin gain spectrum restoration method according to any one of claims 1-8, characterized in that, include: The data processing module is used to acquire complete noise-free Brillouin gain spectrum data and incomplete noise-free Brillouin gain spectrum data, and process the complete noise-free Brillouin gain spectrum data and the incomplete noise-free Brillouin gain spectrum data to obtain complete noise-brillouin gain spectrum data and incomplete noise-brillouin gain spectrum data. The first network construction module is used to construct a variational autoencoder network and a supervised variational autoencoder network, wherein the variational autoencoder network includes a first encoder and a first decoder, and the supervised variational autoencoder network includes a second encoder and a second decoder. The network optimization module is used to optimize the variational autoencoder network using complete noisy Brillouin gain spectrum data, and simultaneously optimize the supervised variational autoencoder network using incomplete noisy Brillouin gain spectrum data. The second network construction module is used to select the optimized second encoder and the first decoder to form a reductive supervised variational autoencoder network. The restoration module is used to process the noisy, incomplete Brillouin gain spectrum using the restoration-supervised variational autoencoder network to obtain complete Brillouin gain spectrum data.
10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the Brillouin gain spectrum restoration method as described in any one of claims 1 to 8.