BOTDA strain extraction method based on RMCHN

Abstract

The invention discloses a BOTDA strain extraction method based on RMCHN, and relates to the technical field of BOTDA strain extraction, and the method comprises the steps: giving a corresponding strain label to a distance-frequency diagram for constructing a training set; training the RMCHN network through the training set so as to enable the RMCHN network to learn a mapping relation between the distance-frequency diagram and the strain; and carrying out strain extraction through the trained RMCHN network. According to the method, the strain extraction of the BOTDA can be realized through the RMCHN network model.

Description

Technical Field

[0001] The present invention relates to the technical field of BOTDA strain extraction, and in particular to a BOTDA strain extraction method based on RMCHN. Background Art

[0002] BOTDA (Brillouin Optical Time Domain Analysis) is a fiber-optic sensing technology based on the principle of Brillouin scattering and is widely used for strain and temperature measurement. Its main application areas include structural health monitoring, earthquake monitoring, and energy transmission safety monitoring. Strain corresponds to the axial deformation of the optical fiber. When measured by the BOTDA system, it can reflect the stretching caused by the external environment at that point on the optical fiber. Due to the long distance of the optical fiber link, the BOTDA system generates a large amount of data during the sensing process. The proposed RMCHN network to implement BOTDA stress extraction technology can be applied to the process of massive data processing, thereby improving the sensing speed of the BOTDA system. At the same time, due to the robustness of the network, the system's tolerance for low signal-to-noise ratio will be further improved, thereby improving sensing accuracy and distance.

[0003] BOTDA technology plays an important role in these fields due to its high sensitivity and long-distance measurement capabilities. It can provide high-precision strain data, helping to improve the safety and reliability of the system. The BOTDA system inputs counter-propagating pulsed pump light and probe light with a certain frequency difference into the system. The two interact with each other in the optical fiber, and the probe light will experience different gains at different modulation frequencies. These gains are detected along the trajectory formed by the optical fiber. The Brillouin gain spectrum (BGS) formed after the gain is detected can be used to obtain the peak value of the Brillouin gain at each measurement position after spectral fitting. The expression of this gain spectrum is as follows:

[0004]

[0005] Among them, g B represents the Brillouin gain coefficient, Δv B represents the Brillouin gain linewidth, v B represents the Brillouin Frequency Shift (BFS) of the system's original state, and v represents the pump-probe frequency difference, also known as the modulation frequency. Due to changes in ambient temperature and strain at the measurement points along the fiber, the BFS at these locations differs from the original BFS. This variable can be expressed as follows:

[0006] v′ B =C T,BFS ΔT+C ε,BFS Δε+vB (2)

[0007] Among them, C T,BFS ≈1.07MHz / ℃, C ε,BFS ≈0.046MHz / με, respectively, represents the temperature and strain sensitivity of the BFS in a standard single-mode fiber. Although similar sensitivities can be found for various optical fibers, due to thermal strain in the fiber coating, the two parameters can in some cases be completely different from the above values. The measured BFS is also known as v' B , which is consistent with v B The corresponding temperature and strain values can be obtained by calculating the difference.

[0008] Recent research in BOTDA data processing has explored the use of advanced deep learning models, such as the Recurrent Modular Convolutional Hypernetwork (RMCHN), to improve the accuracy and efficiency of data analysis. The goal is to extract valid BGS information from the noisy BGS and convert it into strain distribution along the fiber. The RMCHN network establishes a complex, nonlinear mapping relationship between the BGS image s and the strain distribution v. The data generated by BOTDA systems is massive and complex, involving extensive time and frequency domain information, posing a challenge to traditional data processing methods. The RMCHN network, combining the characteristics of recurrent neural networks, modular convolutional networks, and hypernetworks, possesses powerful feature extraction and modeling capabilities, enabling it to process this complex data. This enables faster and more accurate strain distribution measurements. The recurrent neural network portion of the network helps capture temporal relationships in the data, the modular convolutional network effectively extracts local features, and the hypernetwork generates and adjusts model parameters to adapt to varying measurement conditions. Through these mechanisms, the RMCHN network is able to better extract useful information from BOTDA data and improve the accuracy of strain and temperature detection. In summary, applying RMCHN to BOTDA data processing can not only improve the accuracy of data analysis, but also accelerate the response capability of real-time monitoring and early warning systems, providing more reliable data support for key areas such as structural health monitoring, earthquake prediction, and energy pipeline safety. Summary of the Invention

[0009] The embodiment of the present invention provides a BOTDA strain extraction method based on RMCHN, which can realize BOTDA strain extraction through the RMCHN network model.

[0010] In a first aspect, the present invention provides a BOTDA strain extraction method based on RMCHN, comprising:

[0011] Give corresponding strain labels to the distance-frequency graph of the constructed training set;

[0012] The RMCHN network is trained using the training set so that the RMCHN network learns the mapping relationship between the distance-frequency graph and the strain;

[0013] Strain extraction is performed through the trained RMCHN network.

[0014] In some examples, the step of assigning corresponding strain labels to the distance-frequency graph of the constructed training set includes:

[0015] Strain extraction is completed based on the gain spectrum of each measurement point;

[0016] Set strain labels for the images corresponding to each different strain condition.

[0017] In some examples, the Brillouin linewidth variation range is set to 25 MHz to 80 MHz, with a step size of 1 MHz.

[0018] In some examples, a noisy BGS with a signal-to-noise ratio of 10 dB is added to the training set. The BGS are stacked by distance, and the amplitude is attenuated according to the loss coefficient along the optical fiber. The number of pixels in the final distance-frequency shift image is N = frequency sampling interval × distance window length / distance sampling interval.

[0019] In some instances, in the RMCHN network, heterogeneous convolutions are formed by combining convolutions with kernel sizes of 3×3 and 1×1, and dilated convolutions are introduced in FEB. At the same time, transposed convolutions are used to obtain global and local features, thereby reconstructing rough features. Finally, five cascaded 3×3 convolutional layers plus a convolution filter are used to further improve feature accuracy, thereby achieving the effect of outputting the final strain distribution.

[0020] In some instances, the corresponding data label is the center frequency position of the BGS. The strain is calibrated using the gauge coefficient to obtain the corresponding strain force. The BGS and center frequency pair are used as the input and label of the RMCHN network for supervised learning, so that the RMCHN network establishes a mapping relationship between BGS and strain.

[0021] In some instances, a normalized strain method is used in label processing of training data.

[0022] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0023] First, the RMCHN network, through its deep network structure and sophisticated feature extraction capabilities, is able to extract more refined strain information from the gain spectrum acquired by the BOTDA system. Its efficient feature extraction and learning mechanism significantly improves the resolution and accuracy of strain data, thereby reducing the impact of errors and noise during strain measurement. Second, the RMCHN network exhibits strong generalization capabilities and can effectively handle data variations under diverse environments and conditions, resulting in improved robustness and reliability in practical applications. Furthermore, the RMCHN network's training process automatically optimizes parameters, reducing the reliance on manual adjustments in traditional methods and making the system more convenient and efficient. Overall, the application of the RMCHN network in BOTDA strain extraction improves the accuracy, robustness, and automation of data processing, positively impacting the precision and reliability of strain measurements in practical engineering and scientific research. Furthermore, the RMCHN network model typically exhibits excellent speed performance in curve fitting tasks. Benefiting from its parallel processing capabilities, efficient feature extraction, optimized network structure, and automated parameter adjustment, the network is generally faster than traditional methods or other models when performing curve fitting, significantly reducing computation time while maintaining high accuracy. This enables the RMCHN network to achieve fast and accurate responses when processing BOTDA strain data. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0025] Figure 1 is a schematic diagram of a method provided by an embodiment of the present invention;

[0026] Figure 2 Schematic diagram of the Brillouin gain spectrum obtained by the BOTDA system according to an embodiment of the present invention, wherein (a) is the Brillouin gain spectrum of a single point of the optical fiber under test, and (b) is the gain spectrum distribution along the optical fiber under test;

[0027] Figure 3 This is a diagram of the RMCHN network structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0029] In the following description, specific embodiments of the present invention will be described with reference to steps and symbols performed by one or more computers, unless otherwise specified. Therefore, these steps and operations will be mentioned several times as being performed by a computer, and computer execution as referred to herein includes operations by a computer processing unit that represents electronic signals of data in a structured form. This operation converts the data or maintains it at a location in the computer's memory system, which can be reconfigured or otherwise change the operation of the computer in a manner familiar to testers in the field. The data structure in which the data is maintained is a physical location in the memory, which has specific characteristics defined by the data format. However, the principles of the present invention are described in the above text, which does not represent a limitation, and testers in the field will understand that the various steps and operations below can also be implemented in hardware.

[0030] As used herein, the terms "module" or "unit" may be considered software objects executed on the computing system. The various components, modules, engines, and services herein may be considered implementation objects on the computing system. While the devices and methods herein are preferably implemented in software, they may also be implemented in hardware and remain within the scope of protection of the present invention.

[0031] It will be understood by those skilled in the art that, unless expressly stated otherwise, the singular forms "a", "an", and "the" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, "connected" or "coupled" as used herein may include wireless connections or wireless couplings. The term "and / or" used herein includes all or any unit and all combinations of one or more associated listed items.

[0032] The present invention faces the following technical problems in realizing the simultaneous extraction of temperature and strain of BOTDA through RMCHN network:

[0033] 1. The design of the training dataset for the RMCHN network model is crucial. Because the RMCHN network introduces dilated convolution technology, the design of the training dataset directly impacts the model's learning, generalization, and ultimate performance. It requires representativeness, diversity, and high quality. It also requires a balance of various training data types to prevent the model from favoring common categories during prediction. Therefore, the structure of the training dataset for the network model is a technical issue that needs to be addressed.

[0034] 2. Preprocessing the RMCHN network model's input data is a crucial step in ensuring the effectiveness of network training and the stability of the results. For strain extraction tasks, input data requires preprocessing and regularization, and labels during supervised learning require normalization to ensure effective network training. Furthermore, since strain measurement is dependent on the length of the original stretched fiber segment, specific network label design is required. Therefore, the preprocessing of network model input data is a technical issue worthy of further research.

[0035] 3. The impact of the RMCHN network model on strain extraction requires further research. For example, the input layer needs to adapt to the number of data points in the frequency sweep window, the convolution layer needs to continuously adjust the convolution kernel size to match the image characteristics, and the output layer needs to output strain labels in a consistent manner. This work requires continuous trial and error during network training.

[0036] In an embodiment of the present invention, a BOTDA strain extraction method based on RMCHN is provided, such as Figure 1 As shown, the following steps are included:

[0037] S101: giving corresponding strain labels to the distance-frequency graph of the constructed training set;

[0038] S102: training the RMCHN network using the training set, so that the RMCHN network learns the mapping relationship between the distance-frequency graph and the strain;

[0039] S103: Strain extraction is performed using the trained RMCHN network.

[0040] In an embodiment of the present invention, the network training process requires that corresponding strain labels be assigned to the images used to construct the training set. Since the training images are distance-frequency graphs, the features required to complete strain extraction are the peak position, line width, amplitude, noise characteristics, etc. of the gain spectrum at each measurement point. For each different strain condition, corresponding strain labels are assigned to the images. This allows the correct strain label to be assigned to each training image during supervised learning, helping the RMCHN network model learn the mapping relationship between image features and strain, thereby achieving accurate strain extraction. The strain calculation formula is as follows:

[0041]

[0042] Where ΔL represents the elongation of the stretched fiber segment, and L represents the original length of the stretched fiber segment. Substituting this equation into equation (2) calculates the change in BFS for the stretched fiber segment corresponding to a given strain value. After measuring the gain spectrum using BOTDA, the corresponding strain value can be directly obtained through RMCHN processing.

[0043] In the embodiment of the present invention, the application of the RMCHN network to strain extraction in the BOTDA sensing system has the following technical points: First, regarding the setting of the training data set, considering the generalization ability of the trained network model, the training data set should be constructed using simulation data. The network needs to learn the BGS characteristics when variables such as strain and linewidth vary over a large range to achieve accurate temperature extraction. According to formula (1), the Brillouin gain coefficient and phonon lifetime depend on the type of optical fiber to be tested. The impact of strain changes on the BGS spectrum mainly lies in the change of BFS. Corresponding to the range of BFS offset δv, the strain offset Δε is set to a range of 0με to 6000με, with a step size of 50με. The Brillouin linewidth in standard single-mode optical fiber is basically around 30MHz. The Brillouin linewidth in the formula is set to a range of 25MHz to 80MHz, with a step size of 1MHz. To increase the robustness of the network, a noisy BGS with a signal-to-noise ratio of 10dB is added to the training set. These BGSs are then stacked by distance, with their amplitudes attenuated according to the loss coefficient along the fiber. The resulting distance-frequency shift image has a pixel count N = frequency sampling interval × distance window length / distance sampling interval. Specifically, the training data labels require the use of normalized strain to ensure the network outputs accurate results for fibers with varying stretch lengths.

[0044] In the embodiments of the present invention, the RMCHN network model is a modular network. A modular network is a concept in network structure design. Generally, a network model contains two or more submodules. These submodules learn parameters in their own information spaces without interfering with each other. Ultimately, the outputs of each submodule are fused together to achieve the target task. This divides the model into independent components, each of which performs its own function. This concept has been adopted by many research institutions. Modular networks take into account the heterogeneity of information used to understand the same concept, which may come from different fields or perspectives. Mixing these components together without any processing will reduce the quality of learned features.

[0045] In an embodiment of the present invention, before inputting an image into the RMCHN network model, it is usually necessary to preprocess the image. Image preprocessing can increase the diversity of the data, thereby helping the model to better generalize to new data. Image normalization, as a general image preprocessing technology, can make the input image data have a uniform numerical range and distribution, which helps to avoid deviations between different data features and accelerate the convergence process of the model. Normalizing the image pixel values to the range of [0,1] is a general image preprocessing technology, that is, each pixel value of the image is divided by 255 (that is, the maximum pixel value of the image) and scaled to between 0 and 1. This method is suitable for most cases. At the same time, since the label variation range in the training data is large, considering the ability of the network model to migrate, the labels during training also need to be normalized, which can help the model better utilize the data, reduce overfitting or underfitting problems, and improve the model's adaptability and generalization capabilities to real-world data.

[0046] Figure 2 (a) shows the Brillouin gain spectrum measured using BOTDA. The BGS exhibits a distinct Lorentzian lineshape, and conventional curve fitting can be used to obtain the peak frequency of the gain spectrum and infer the strain value at that point in the fiber under test. Figure 2 (b) shows the distribution of BGS along the optical fiber. This image is obtained by stacking the Brillouin gain spectra over distance. It can be seen that when stress is applied to a certain part of the optical fiber, the position of the BGS gain peak will change. This change can be used as an image feature to be learned by the RMCHN network, thereby realizing the function of strain extraction.

[0047] Figure 3 The network architecture and flow chart for obtaining the stress distribution along the optical fiber using RMCHN to process Brillouin gain spectrum data obtained from BOTDA measurements are presented. The goal is to combine convolutions with kernel sizes of 3×3 and 1×1 to form heterogeneous convolutions, preventing information loss while reducing the number of parameters. To increase the receptive field without sacrificing spatial resolution, dilated convolutions are introduced into FEB. Dilated convolutions can, to a certain extent, expand the receptive field without changing the kernel size. In other words, when feature maps are of equal size, dilated convolutions can achieve a larger receptive field and denser data. Furthermore, features are integrated to avoid the loss of shallow information with increasing network depth. Transposed convolutions are used to obtain global and local features, thereby reconstructing coarse features. Coarse features are shallow, inaccurate, and incomplete in reflecting the underlying information. Finally, five cascaded 3×3 convolutional layers plus a convolutional filter further improve feature accuracy, resulting in the final output of the strain distribution.

[0048] In an embodiment of the present invention, the cascaded module convolutional neural network, with its deep structure and complex feature extraction technology, is able to parse more detailed strain data from the gain spectrum captured by the distributed Brillouin optical time-domain analysis sensing system. During the training phase, it is first necessary to construct a training data set. In order to ensure that the trained network model has good generalization capabilities, these data sets should be generated through simulation data. The model must be able to identify the Brillouin gain spectrum (BGS) characteristics when parameters such as strain and line width change significantly in order to achieve accurate strain measurement. According to formula (1), the Brillouin gain coefficient and phonon lifetime are related to the type of optical fiber being measured, and the strain change mainly affects the frequency shift (BFS) of the BGS spectrum line. Figure 2 (a) shows the envelope shape of a single BGS, and Figure 2 (b) shows the arrangement of BGS along the optical fiber. It can be seen that the BGS exhibits a significant peak displacement in the fiber segment where strain is applied, and the displacement is proportional to the stress applied to the fiber at that location. Therefore, the strain offset Δε is set to range from 0 microstrain to 6000 microstrain, with a step size of 50 microstrain. For standard single-mode optical fiber, the Brillouin linewidth is typically around 30 MHz, and the linewidth variation range is set to range from 25 MHz to 80 MHz, with a step size of 1 MHz. To enhance the robustness of the network, the training set should also include noisy BGS samples under different signal-to-noise ratios. These BGS samples will be stacked based on distance and attenuated in amplitude based on the loss coefficient along the optical fiber. The number of pixels N in the resulting distance-frequency shift image can be calculated from the frequency sampling interval, the distance window length, and the distance sampling interval. The resulting dataset contains a total of 121 × 56 × 4 = 27,104 distinct BGSs, with the corresponding strain labels being the center frequency positions of the BGSs. Strain calibration using the gauge factor allows the corresponding strain force to be derived. Using the BGS and center frequency pairs as the network input and label for supervised learning, the network establishes a mapping between BGSs and strains, transforming input BGSs into output strains. Normalized strain is employed in the labeling of the training data to ensure the network accurately outputs results for fibers with varying stretch lengths. This means that, regardless of fiber length, the network can correctly identify and output strain measurements by learning the normalized strain signature. The root mean square error (RMS) of the test set extraction results is used as a measure of network performance, with training concluded when it is below 50 microstrain. Figure 3 Shows the specific structure of the network.

[0049] In the embodiment of the present invention, since the actual sensing system has unpredictable noise, in order to verify the performance of the trained network, it is necessary to build a distributed Brillouin optical time domain analysis sensing system. In order to verify the performance of the network under the condition of long sensing distance, a 50km optical fiber to be tested is selected, and the last 20m is wound down and wrapped around a stress disk. By adjusting the scale of the stress disk, different stresses are applied to the optical fiber, and 0, 2000, 4000, 6000, and 8000 microstrains are applied respectively. The BGS distribution along the optical fiber is measured, which is as follows: Figure 2 (b) These data images are preprocessed and then fed into the network. The network processes the BGS at each point to obtain the corresponding stress value. By comparing the output stress value with the actual value, the root mean square error (RMS) of the strain extraction under different conditions can be calculated.

[0050] The above is a detailed introduction to a BOTDA strain extraction method based on RMCHN provided in an embodiment of the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only intended to help understand the method and core concept of the present invention. At the same time, for those skilled in the art, according to the concept of the present invention, there may be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A BOTDA strain extraction method based on RMCHN, characterized in that: include: Give corresponding strain labels to the distance-frequency graph of the constructed training set; The RMCHN network is trained using the training set so that the RMCHN network learns the mapping relationship between the distance-frequency graph and the strain; Strain extraction is performed through the trained RMCHN network.

2. The method according to claim 1, characterized in that The step of giving corresponding strain labels to the distance-frequency graph of the training set includes: Strain extraction is completed based on the gain spectrum of each measurement point; Set strain labels for the images corresponding to each different strain condition.

3. The method according to claim 2, characterized in that The strain offset Δε was set to range from 0 microstrain to 6000 microstrain with a step size of 50 microstrain, and the Brillouin linewidth was set to range from 25 MHz to 80 MHz with a step size of 1 MHz.

4. The method according to claim 3, characterized in that A noisy BGS with a signal-to-noise ratio of 10 dB was added to the training set. The BGS were stacked by distance, and the amplitude was attenuated according to the loss coefficient along the optical fiber. The number of pixels in the final distance-frequency shift image was N = frequency sampling interval × distance window length / distance sampling interval.

5. The method according to claim 4, characterized in that In the RMCHN network, heterogeneous convolution is formed by combining convolutions with kernel sizes of 3×3 and 1×1, and dilated convolution is introduced in FEB. At the same time, transposed convolution is used to obtain global and local features, thereby reconstructing rough features. Finally, five cascaded 3×3 convolution layers plus a convolution filter are used to further improve feature accuracy, thereby achieving the effect of finally outputting strain distribution.

6. The method according to claim 5, characterized in that The corresponding data label is the center frequency position of BGS. The strain is calibrated by the gauge coefficient to obtain the corresponding strain force. The BGS and center frequency are used as the input and label of the RMCHN network for supervised learning, so that the RMCHN network can establish a mapping relationship between BGS and strain.

7. The method according to claim 6, characterized in that In the label processing of training data, the normalized strain method is adopted.

Images