Fiber grating stress sensor demodulation model selection method based on calibration data

Abstract

The invention provides a fiber bragg grating (FBG) stress sensor demodulation model selection method based on calibration data, and aims to improve the test efficiency and precision of an FBG sensor. The invention belongs to the field of fiber bragg grating sensing, and is particularly suitable for the inspection process of fiber bragg grating sensor manufacturing. Compared with a traditional method, the method has the remarkable advantages that high efficiency is achieved, the classification and screening process is simplified through fast processing of the neural network, and dependence on complex experimental equipment is reduced; the method is simple and accurate, high-precision calibration of parameters of the FBG stress sensor is realized by optimizing a neural network model, and selection of an accurate demodulation model is provided; and the calibration time and the experiment cost are reduced.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology, fiber optic sensor applications, and the application technology of neural network in fiber grating sensors. Specifically, it is a method for selecting a demodulation model of a fiber grating stress sensor based on calibration data, aiming to achieve high-precision demodulation and sensor classification. Background Art

[0002] In the context of continuous economic development and rapid increase in information capacity, fiber Bragg grating sensing technology has been widely used and valued, becoming one of the hot research directions at home and abroad. Fiber Bragg Grating (FBG) sensor is an advanced sensing technology based on the principle of optical interference and the characteristics of optical fiber. It is widely used in aerospace, petrochemical, building damage monitoring and other fields due to its small size, corrosion resistance, and low sensitivity to electromagnetic interference.

[0003] During the manufacturing process of fiber Bragg grating sensors, although the same process and materials are usually used, the same batch of FBG sensors may have differences due to manufacturing process variables, material property differences, or manual operation effects. These differences may arise from slight process fluctuations during the grating stretching process, such as changes in tension or temperature, and slight differences in refractive index between different batches of material suppliers. In addition, slight differences in manual operation accuracy and position during the manufacturing process may also affect the structure and performance of the final grating. Although these differences are generally acceptable in most applications, in some applications where high accuracy and consistency are required, strict quality control and screening measures are required to ensure the performance stability and reliability of the fiber Bragg grating.

[0004] Traditional fiber Bragg grating sensor classification and demodulation methods are usually based on physical models and signal processing techniques, such as edge filtering and matched filtering. These methods rely on mathematical modeling of spectral response curves and complex signal processing algorithms. Although they have a certain degree of accuracy in theory, they face many limitations. For example, they require a lot of prior knowledge, are sensitive to environmental conditions, are difficult to adapt to complex and dynamic strain scenarios, and have challenges in real-time and automation. The limitations of traditional methods restrict the application and performance of fiber Bragg grating sensors in complex engineering environments.

[0005] In recent years, the application of neural networks in the field of fiber Bragg grating sensing has gradually increased, and its role and advantages have been widely recognized, mainly reflected in the use of its powerful pattern recognition and feature extraction capabilities. Neural networks use their powerful data-driven learning capabilities to automatically learn complex spectral response patterns and nonlinear characteristics from a large amount of experimental data, and can more accurately demodulate the signals of fiber Bragg grating sensors than traditional methods. Its ability to adapt to different environmental conditions and process real-time data has shown important application potential in industrial control, structural health monitoring and other fields. In addition, the automation characteristics of neural networks reduce human errors and operating costs, and improve the stability and reliability of demodulation. Overall, the application of neural networks in fiber Bragg grating sensing technology continues to expand, and has opened up new possibilities for improving sensor demodulation accuracy, real-time performance and application scope.

[0006] The linear fitting model and the neural network model each have their own advantages and disadvantages. The advantage of the linear fitting model is that it has a simple structure, a small amount of calculation, and the model parameters are easy to obtain, so it is relatively convenient in implementation and calculation. However, the linear model has poor adaptability and cannot be effectively applied to complex nonlinear scenarios, which limits its scope of use. In contrast, the neural network model is relatively complex, has a large amount of calculation, the model parameters are not easy to obtain, and a large amount of data is required for training. However, the neural network model has strong adaptability, can handle various nonlinear scenarios, and significantly improves the accuracy of the model. The present invention combines the advantages of these two models, and realizes more accurate and efficient parameter calibration through the simplicity of the linear fitting model and the strong adaptability of the neural network model. Summary of the invention

[0007] The purpose of the present invention is to provide a method for selecting a demodulation model of a fiber Bragg grating stress sensor based on calibration data, which can effectively classify the performance of a batch of newly produced fiber Bragg grating sensors. The technical solution of the present invention is as follows:

[0008] In the first step, a uniform linearity test is performed on all newly manufactured K fiber grating sensors (103). A stress experiment test is performed on each FBG sensor under L stress variables, with the stress range uniformly changing within [W1, W2], and the spectral data of the 1*N sequence {P1(1, N), P2(1, N), ...P l (1, N)...P L (1, N)}, P L (1, N) represents the spectral sequence measured once under the Lth stress condition, N represents the Nth point scanned by the spectrometer, the spectral range is [f1, f2], the sampling interval is Δf, and the number of sampling points is

[0009] Furthermore, the conventional linear model centroid method is used for the goodness of fit test. The centroid method determines the position of the central wavelength by calculating the weighted average position of the reflection intensity distribution. The central wavelength is determined as the centroid of the FBG reflection spectrum. The formula is: p l (λ n ) is the lth stress, the wavelength λ n where a spectrum can be expressed as P l (1, N) = [p l (λ1)p l (λ2)...p l (λ n )...p l (λ N )], the spectra at different stresses form a spectrum sequence {P1(1, N), P2(1, N), ...P l (1, N)...P L (1, N)}, this spectrum sequence is input into the linear model, and the demodulation results in the following wavelength sequence represents the estimated value of the central wavelength under the lth stress condition, and linear represents the linear model method used;

[0010] Furthermore, the linear fit is calculated, and the formula is: The linear fit R 2 A FBG sensor that reaches 0.99 and above is classified as being ready for direct use.

[0011] Furthermore, for the remaining sensors that do not meet the standards, it is necessary to increase the test volume and use L stress ranges to test the remaining K - A FBG sensors are subjected to M repetitive experiments under the same stress conditions to obtain enough data for 1D CNN (one-dimensional convolution) neural network optimization training; the total data volume is (KA)*L*M*N.

[0012] Furthermore, for each of the KA FBG sensors, a spectral sequence P is repeated M times in the L stress ranges. l (M, N), 1≤l≤L, the lth stress W l =W1+(W2-W1) / ( L -1), W1 and W2 are the starting and ending stress values ​​respectively; the total data volume of KA FBG sensors is L*M*N, among which the spectral data under the lth stress condition is P l (M, N), expressed as:

[0013]

[0014] Among them, P l In (M, N), l represents the lth stress condition, M represents the number of repeated data collection, N represents the maximum sampling point of spectrum collection, m, n represent the sequence number, 1≤m≤M, 1≤n≤N; p m (λ n ) indicates the mth acquisition at wavelength λ n The reflection intensity at .

[0015] Furthermore, the centroid method is used to calculate the matrix P l The resonant wavelength of each spectrum in (M, N) is obtained to obtain the resonant wavelength sequence of M spectra under the lth stress condition represents the resonant wavelength of the spectrum data of the Mth test under the lth stress condition, and the average of these resonant wavelengths is obtained The resonance wavelength sequence under L stress conditions Perform linear fitting, and the wavelength sequence after fitting It is used to supervise the learning of the neural network model; through the neural network training, the B FBG sensors with the best results reaching RMSE < 0.01 are classified as the ones that can be trained by the neural network.

[0016] Furthermore, after the spectral data under the stress condition changes are trained to obtain the optimal demodulation result, the optimized 1D CNN neural network model parameters of B FBG sensors that meet the conditions, including the number x and size y of the convolution kernels in the convolution layer (302), the stride z; the type o (such as maximum pooling or average pooling) and the window size v of the pooling layer (303); the number r of neurons in the fully connected layer (304); the type p of the activation function (such as ReLU, Sigmoid); and the learning rate j of the optimizer, are recorded in the database for direct call in subsequent engineering applications.

[0017] Furthermore, the remaining KAB FBG sensors with poor neural network training results are classified as unusable and discarded.

[0018] Furthermore, the experimental system of the method includes the following components:

[0019] a) An erbium-doped fiber amplifier (101), as a broadband light source, is used to amplify optical signals so as to compensate for the attenuation of optical signals in long-distance optical fiber transmission (Erbium-Doped Fiber Amplifier, EDFA).

[0020] b) a circulator (102), used for controlling the transmission direction of the optical signal, thereby realizing separation and guidance of the optical signal; the circulator (102) comprises three ports, port 1 is connected to the wide spectrum light source EDFA, port 2 is connected to the FBG sensor, and port 3 is connected to the spectrometer (105).

[0021] c) Fiber Bragg grating sensor (103), mainly used to measure changes in physical quantities, such as strain, temperature, pressure, etc.; its working principle is based on the reflection characteristics of the Bragg grating in the optical fiber to light of a specific wavelength;

[0022] d) The two-end clamping structure experimental platform is designed to accurately apply axial stress and can provide a minimum variable unit of 0.01mm to test the performance of FBG sensors.

[0023] e) A spectrometer (105) for detecting and outputting the spectrum of the optical signal reflected or transmitted by the fiber grating.

[0024] f) A neural network analysis module, whose main structure is a 1D CNN structure model, is used to analyze a large amount of spectral data output by the spectrometer (105) and demodulate the relationship between stress and central wavelength from the complex spectral data.

[0025] Furthermore, the two-end clamping structure experimental platform includes the following structure:

[0026] The platform comprises a micro-motion platform 1 (201), a micro-motion platform 2 (202), a support platform (203), and a clamp (204); two ends of the FBG sensor are respectively fixed to the micro-motion platform 1 (201) and the micro-motion platform 2 (202) by the clamp (204); the support platform (203) fixes the FBG sensor to the same height; the FBG sensor generates strain by moving the micro-motion platform 2 (202), so that a corresponding relationship can be generated according to the displacement of the micro-motion platform 2 (202) and the drift of the central wavelength of the optical fiber grating reflection spectrum.

[0027] Furthermore, the algorithm of the neural network analysis module includes the following parts:

[0028] The internal parameters of the neural network are continuously debugged to achieve the optimal effect, including the following parameters: the number x and size y of convolution kernels, stride z; the type o (such as maximum pooling or average pooling) and window size v of the pooling layer (303); the number r of neurons in the fully connected layer (304); the type p of the activation function (such as ReLU, Sigmoid); and the learning rate j of the optimizer; by continuously optimizing the parameters, after demodulation, the following sequence is obtained,

[0029]

[0030] Further, represents the estimated value of the central wavelength under the lth stress condition, and cnn represents the convolutional neural network model used.

[0031] Furthermore, the output of the neural network is the root mean square error (RMSE) after training with different parameters. RMSE is used to measure the difference between the model prediction value and the actual value. Its formula is: By calculating the demodulation results W of (KA) fiber Bragg grating sensors (103) under different parameters cnn(L) , and the input learning sequence W l The minimum RMSE value obtained by training KA sensors is recorded as {RMSE1, RMSE2, ..., RMSEK K-A}, the B FBG sensors with the best results reaching RMSE < 0.01 are classified as the ones that can be trained with neural networks.

[0032] Further, the input layer (301) in the neural network is based on the spectral data under each stress condition (each spectrum has The fiber Bragg grating sensor (103) data is first read and standardized, and then divided into input (x) and output (y). The data set is divided into 90% training set and 10% test set, and managed by custom data set class and data loader. A suitable computing device (CPU or GPU) is selected for training, and the parameter calibration of the sensor is optimized through a neural network, ultimately improving the accuracy of the fiber Bragg grating stress sensor and the reliability of engineering applications.

[0033] Furthermore, the one-dimensional convolution is used in the neural network, and its operation is shown in the formula, where x(t) is the input signal, w(s) is the convolution kernel, and the convolution kernel size is assumed to be n.

[0034]

[0035] Furthermore, to extract features from the input data through convolution operations, it is necessary to apply activation functions, such as ReLU, Sigmoid, Tanh, etc., to perform nonlinear transformations on the convolution output. The advantage of the ReLU function is that it increases the sparsity of the network, has good generalization performance and small computational complexity. The disadvantage is that sparse processing will result in the inability to learn effective features, so we choose the ReLU function.

[0036] ReLU(x)=max(0,x)

[0037] Furthermore, after the activation function, a pooling operation (such as maximum pooling or average pooling) may be performed to reduce the spatial size of the feature map and reduce the model complexity. Assuming that the pooling kernel size is 2, the expressions of one-dimensional maximum pooling and one-dimensional average pooling are as follows. We choose maximum pooling in the network because maximum pooling focuses more on retaining the significant part of the feature and ignoring other information.

[0038] P(t)=max[y′(2t),y′(2t+1)]

[0039]

[0040] Furthermore, the fully connected layer (304) is the last layer of the CNN, which is responsible for organizing the feature information after pooling, extracting and saving useful information. Among them, P is the output of the previous layer, W is the weight of the fully connected layer (304), and B is the bias. The process of implementing the fully connected layer (304) embodies linear characteristics because it performs calculations by linearly combining the input and weights.

[0041] Y=W.P+B

[0042] Furthermore, under this structured experimental platform, users can easily replace and test different FBG sensors, whether they are FBG sensors produced in the same batch or FBG sensors with different packages. These FBG sensors will inevitably have slight differences in the production process, such as packaging methods, material properties or manufacturing processes. This feature enables the system to effectively deal with slight differences between sensors and ensure accurate and consistent calibration of sensors with various packaging specifications. Through this flexibility, users can select and apply the most suitable sensors in different stress measurement and monitoring scenarios, thereby improving the applicability and practicality of the system.

[0043] Furthermore, in the database, key physical parameters such as stress value, wavelength movement, fit and root mean square error are recorded and stored in the database; in addition, the optimized neural network model parameters are also saved for direct call and application in subsequent engineering applications; users can use the query function of the database to instantly obtain the required parameter data according to the sensor identification or specific test conditions, so as to use this information in engineering applications and scientific research experiments to optimize sensor performance and improve its application effect and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of the overall process of a method for selecting a demodulation model of a fiber Bragg grating stress sensor based on calibration data according to the present invention;

[0045] Figure 2 A schematic diagram of the structure of the experimental platform provided by the present invention;

[0046] Figure 3 A schematic diagram of an experimental platform with a clamping structure at both ends provided by the present invention;

[0047] Figure 4 A schematic diagram of the neural network structure provided by the present invention; DETAILED DESCRIPTION

[0048] The technical solution of the present invention will be further described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0049] like Figure 1 As shown, the present invention provides a method for selecting a demodulation model of a fiber Bragg grating stress sensor based on calibration data, and the process is as follows:

[0050] In the first step, a uniform linearity test is performed on all newly manufactured K fiber grating sensors (103). A stress experiment test is performed on each FBG sensor under L stress variables, with the stress range uniformly changing within [W1, W2], and the spectral data of the 1*N sequence {P1(1, N), P2(1, N), ...P l (1, N)...P L (1, N)}, P L (1, N) represents the spectral sequence measured once under the Lth stress condition, N represents the Nth point scanned by the spectrometer, the spectral range is [f1, f2], the sampling interval is Δf, and the number of sampling points is

[0051] More specifically, the conventional linear model centroid method is used for the goodness of fit test. The centroid method determines the position of the central wavelength by calculating the weighted average position of the reflection intensity distribution. The central wavelength is determined as the centroid of the FBG reflection spectrum. The formula is: p l (λ n ) is the lth stress, the wavelength λ n where a spectrum can be expressed as Pl(1, N) = [p l (λ1)p l (λ2)...p l (λ n )...P l (λ N )], the spectra at different stresses form a spectrum sequence {P1(1, N), P2(1, N), ...P l (1, N)...P L (1, N)}, this spectrum sequence is input into the linear model, and the demodulation results in the following wavelength sequence represents the estimated value of the central wavelength under the lth stress condition, and linear represents the linear model method used;

[0052] More specifically, the linear fit is calculated as The linear fit R 2 A FBG sensor that reaches 0.99 and above is classified as being ready for direct use.

[0053] The fourth step is to increase the test volume for the remaining sensors that do not meet the standards. Use L stress ranges to conduct M repetitive experiments on the remaining KA FBG sensors under the same stress conditions to obtain enough data for 1DCNN (one-dimensional convolutional) neural network optimization training; the total data volume is (KA)*L*M*N.

[0054] More specifically, for each of the KA FBG sensors, a spectral sequence P is repeated M times in the L stress ranges. l (M, N), 1≤l≤L, the lth stress W l =W1+(W2-W1) / (L-1), W1 and W2 are the starting and ending stress values ​​respectively; the total data volume of KA FBG sensors is L*M*N, among which the spectral data under the lth stress condition is P l (M, N), expressed as:

[0055]

[0056] More specifically, P l In (M, N), l represents the lth stress condition, M represents the number of repeated data collection, N represents the maximum sampling point of spectrum collection, m, n represent the sequence number, 1≤m≤M, 1≤n≤N; p m (λ n ) indicates the mth acquisition at wavelength λ n The reflection intensity at .

[0057] More specifically, the centroid method is used to calculate the matrix P l The resonant wavelength of each spectrum in (M, N) is obtained to obtain the resonant wavelength sequence of M spectra under the lth stress condition represents the resonant wavelength of the spectrum data of the Mth test under the lth stress condition, and the average of these resonant wavelengths is obtained The resonance wavelength sequence under L stress conditions Perform linear fitting, and the wavelength sequence after fitting It is used to supervise the learning of the neural network model; through the neural network training, the B FBG sensors with the best results reaching RMSE < 0.01 are classified as the ones that can be trained by the neural network.

[0058] More specifically, after the spectral data under stress conditions are trained to obtain the optimal demodulation result, the optimized 1D CNN neural network model parameters of B FBG sensors that meet the conditions, including the number x and size y of the convolution kernels in the convolution layer (302), the stride z; the type o (such as maximum pooling or average pooling) and the window size v of the pooling layer (303); the number r of neurons in the fully connected layer (304); the type p of the activation function (such as ReLU, Sigmoid); and the learning rate j of the optimizer, are recorded in a database for direct call in subsequent engineering applications.

[0059] More specifically, the remaining KAB FBG sensors with poor neural network training results are classified as unusable and discarded.

[0060] More specifically, an erbium-doped fiber amplifier (101), as a broadband light source, is used to amplify optical signals so as to compensate for the attenuation of optical signals in long-distance optical fiber transmission (Erbium-Doped Fiber Amplifier, EDFA).

[0061] More specifically, the circulator (102) is used to control the transmission direction of the optical signal, thereby realizing the separation and guidance of the optical signal; the circulator (102) comprises three ports, port 1 is connected to the wide spectrum light source EDFA, port 2 is connected to the FBG sensor, and port 3 is connected to the spectrometer (105).

[0062] More specifically, the fiber Bragg grating sensor (103) is mainly used to measure changes in physical quantities, such as strain, temperature, pressure, etc.; its working principle is based on the reflection characteristics of the Bragg grating in the optical fiber to light of a specific wavelength.

[0063] More specifically, the two-end clamping structure experimental platform (104) is designed to accurately apply axial stress and can provide a minimum variable unit of 0.01 mm to test the performance of the FBG sensor.

[0064] More specifically, a spectrometer (105) for detecting and outputting a spectrum of an optical signal reflected or transmitted by a fiber grating;

[0065] More specifically, the neural network analysis module (106), whose main structure is a 1D CNN structural model, is used to analyze a large amount of spectral data output by the spectrometer (105) and demodulate the relationship between stress and central wavelength from the complex spectral data.

[0066] More specifically, the two-end clamping structure experimental platform (104) comprises the following structure: the platform comprises a micro-motion platform 1 (201), a micro-motion platform 2 (202), a support platform (203), and a clamp (204); the two ends of the FBG sensor are respectively fixed to the micro-motion platform 1 (201) and the micro-motion platform 2 (202) by the clamp (204); the support platform (203) fixes the FBG sensor to the same height; by moving the micro-motion platform 2 (202), the FBG sensor generates strain, so that a corresponding relationship can be generated according to the displacement of the micro-motion platform 2 (202) and the drift of the central wavelength of the fiber grating reflection spectrum.

[0067] More specifically, the internal parameters of the neural network are continuously debugged to achieve the optimal effect, including the following parameters: the number x and size y of the convolution kernel, the stride z; the type o (such as maximum pooling or average pooling) and window size v of the pooling layer (303); the number r of neurons in the fully connected layer (304); the type p of the activation function (such as ReLU, Sigmoid); and the learning rate j of the optimizer; by continuously optimizing the parameters, after demodulation, the following sequence is obtained,

[0068]

[0069] More specifically, represents the estimated value of the central wavelength under the lth stress condition, and cnn represents the convolutional neural network model used.

[0070] More specifically, the output of the neural network is the root mean square error (RMSE) after training with different parameters. RMSE is used to measure the difference between the model prediction value and the actual value. Its formula is: By calculating the demodulation results W of (KA) fiber Bragg grating sensors (103) under different parameters cnn(L) , and the input learning sequence W l (M), the minimum RMSE value obtained by training KA sensors is recorded as {RMSE1, RMSE2, ..., RMSE K-A}, the B FBG sensors with the best results reaching RMSE < 0.01 are classified as the ones that can be trained with neural networks.

[0071] More specifically, the input layer (301) in the neural network is based on the spectral data under each stress condition (each spectrum has The fiber Bragg grating sensor (103) data is first read and standardized, and then divided into input (x) and output (y). The data set is divided into 90% training set and 10% test set, and managed by custom data set class and data loader. A suitable computing device (CPU or GPU) is selected for training, and the parameter calibration of the sensor is optimized through a neural network, ultimately improving the accuracy of the fiber Bragg grating stress sensor and the reliability of engineering applications.

[0072] Furthermore, the one-dimensional convolution is used in the neural network, and its operation is shown in the formula, where x(t) is the input signal, w(s) is the convolution kernel, and the convolution kernel size is assumed to be n.

[0073]

[0074] More specifically, to extract features from input data through convolution operations, it is necessary to apply activation functions, such as ReLU, Sigmoid, Tanh, etc., to perform nonlinear transformations on the convolution output. The advantage of the ReLU function is that it increases the sparsity of the network, has good generalization performance and small computational complexity. The disadvantage is that sparse processing will result in the inability to learn effective features, so we choose the ReLU function.

[0075] ReLU(x)=max(0,x)

[0076] More specifically, after the activation function, a pooling operation (such as maximum pooling or average pooling) may be performed to reduce the spatial size of the feature map and reduce the model complexity. Assuming the pooling kernel size is 2, the expressions of one-dimensional maximum pooling and one-dimensional average pooling are as follows. We choose maximum pooling in the network because maximum pooling focuses more on retaining the salient parts of the features while ignoring other information.

[0077] P(t)=max[y(2t),y(2t+1)]

[0078]

[0079] More specifically, the fully connected layer (304) is the last layer of the CNN, which is responsible for organizing the feature information after pooling, extracting and saving useful information. Among them, P is the output of the previous layer, W is the weight of the fully connected layer (304), and B is the bias. The process of implementing the fully connected layer (304) embodies linear characteristics because it performs calculations by linearly combining the input and weights.

[0080] Y=W.P+B

[0081] More specifically, under this structured experimental platform, users can easily replace and test different FBG sensors, whether they are FBG sensors produced in the same batch or FBG sensors with different packages. These FBG sensors will inevitably have slight differences in the production process, such as packaging methods, material properties or manufacturing processes. This feature enables the system to effectively deal with slight differences between sensors and ensure accurate and consistent calibration of sensors with various packaging specifications. Through this flexibility, users can select and apply the most suitable sensors in different stress measurement and monitoring scenarios, thereby improving the applicability and practicality of the system.

[0082] More specifically, in the database, key physical parameters such as stress value, wavelength movement, fit and root mean square error are recorded and stored in the database; in addition, the optimized neural network model parameters are also saved for direct call and application in subsequent engineering applications; users can use the database query function to instantly obtain the required parameter data according to the sensor identification or specific test conditions, so as to use this information to optimize sensor performance in engineering applications and scientific research experiments, and improve its application effect and accuracy.

[0083] This invention significantly improves the accuracy and reliability of fiber Bragg grating sensors in stress measurement by combining traditional experimental tests with modern machine learning technology. This method can not only effectively screen, classify and optimize sensor performance, but also record the optimized neural network model parameters of the FBG sensor into the database, providing convenient and reliable technical support for accurate stress measurement in engineering applications, and is expected to promote the widespread application and further development of fiber Bragg grating technology in practical applications.

Claims

1. A method for selecting a demodulation model for a fiber Bragg grating stress sensor based on calibration data, characterized in that: The method comprises the following steps: In the first step, a uniform linearity test is performed on all newly manufactured K fiber grating sensors (103). A stress experiment test is performed on each FBG sensor under L stress variables, with the stress range uniformly changing within [W1, W2], and the spectral data of the 1*N sequence {P1(1, N), P2(1, N), ...P l (1,N)...P L (1,N)}, P L (1, N) represents the spectral sequence measured once under the Lth stress condition, N represents the Nth point scanned by the spectrometer, the spectral range is [f1, f2], the sampling interval is Δf, and the number of sampling points is In the second step, the traditional linear model centroid method is used to perform a goodness of fit test. The centroid method determines the position of the central wavelength by calculating the weighted average position of the reflection intensity distribution. The central wavelength is determined as the centroid of the FBG reflection spectrum. The formula is: p l (λ n ) is the lth stress, the wavelength λ n where a spectrum can be expressed as P l (1,N)=[p l (λ1)p l (λ2)...p l (λ n )...p l (λ N )], the spectra at different stresses form a spectrum sequence {P1(1, N), P2(1, N), ...P l (1, N)...P L (1, N)}, this spectrum sequence is input into the linear model, and the demodulation results in the following wavelength sequence represents the estimated value of the central wavelength under the lth stress condition, and linear represents the linear model method used; The third step is to calculate the linear fit, the formula is: The linear fit R 2 A FBG sensors that reach 0.99 and above are classified as being directly usable; The fourth step is to increase the test volume for the remaining sensors that do not meet the standards. Perform M repetitive experiments on the remaining KA FBG sensors under the same stress conditions using L stress ranges to obtain enough data for 1D CNN (one-dimensional convolution) neural network optimization training; the total data volume is (KA)*L*M*N; In the fifth step, each of the KA FBG sensors is trained by repeating the spectral sequence P M times in the L stress ranges. l (M, N), 1≤l≤L, the lth stress W l =W1+(W2-W1) / (L-1), W1 and W2 are the starting and ending stress values ​​respectively; the total data volume of KA FBG sensors is L*M*N, among which the spectral data under the lth stress condition is P l (M, N), expressed as: Among them, P l In (M, N), l represents the lth stress condition, M represents the number of repeated data collection, N represents the maximum sampling point of spectrum collection, m, n represent the sequence number, 1≤m≤M, 1≤n≤N; p m (λ n ) indicates the mth acquisition at wavelength λ n The reflection intensity at Step 6: Use the centroid method to calculate P in the matrix l The resonant wavelength of each spectrum in (M, N) is obtained to obtain the resonant wavelength sequence of M spectra under the lth stress condition represents the resonant wavelength of the spectrum data of the Mth test under the lth stress condition, and the average of these resonant wavelengths is obtained The resonance wavelength sequence under L stress conditions Perform linear fitting, and the wavelength sequence after fitting Used to supervise the learning of the neural network model; through neural network training, the B FBG sensors with the best results reaching RMSE < 0.01 are classified as the ones that can be trained by the neural network; Step 7: After the spectral data under stress conditions are trained to obtain the optimal demodulation result, the optimized 1D CNN neural network model parameters of B FBG sensors that meet the conditions, including the number x and size y of the convolution kernels in the convolution layer (302), the stride z; the type o (such as maximum pooling or average pooling) and the window size v of the pooling layer (303); the number r of neurons in the fully connected layer (304); the type p of the activation function (such as ReLU, Sigmoid); and the learning rate j of the optimizer, are recorded in the database for direct call in subsequent engineering applications; In the eighth step, the remaining KAB FBG sensors with poor neural network training results are classified as unusable and discarded.

2. The method for selecting a demodulation model for a fiber Bragg grating stress sensor based on calibration data according to claim 1, wherein the experimental system of the method comprises the following components: An erbium-doped fiber amplifier (101), as a broadband light source, is used to amplify optical signals so as to compensate for the attenuation of optical signals in long-distance optical fiber transmission (Erbium-Doped Fiber Amplifier, EDFA); The circulator (102) is used to control the transmission direction of the optical signal, thereby realizing separation and guidance of the optical signal; the circulator (102) comprises three ports, port 1 is connected to the wide spectrum light source EDFA, port 2 is connected to the FBG sensor, and port 3 is connected to the spectrometer (105); The fiber Bragg grating sensor (103) is mainly used to measure changes in physical quantities, such as strain, temperature, pressure, etc. Its working principle is based on the reflection characteristics of the Bragg grating in the optical fiber to light of a specific wavelength; The two-end clamping structure experimental platform (104) is designed to accurately apply axial stress and can provide a minimum variable unit of 0.01 mm to test the performance of the FBG sensor; A spectrometer (105) for detecting and outputting a spectrum of an optical signal reflected or transmitted by the fiber grating; The neural network analysis module (106) mainly has a 1D CNN structure model and is used to analyze a large amount of spectral data output by the spectrometer (105) and demodulate the relationship between stress and central wavelength from the complex spectral data.

3. The method for selecting a demodulation model for a fiber Bragg grating stress sensor based on calibration data according to claim 1, characterized in that: The two-end clamping structure experimental platform (104) comprises the following structure: The platform comprises a micro-motion platform 1 (201), a micro-motion platform 2 (202), a support platform (203), and a clamp (204); two ends of the FBG sensor are respectively fixed to the micro-motion platform 1 (201) and the micro-motion platform 2 (202) by the clamp (204); the support platform (203) fixes the FBG sensor to the same height; the FBG sensor generates strain by moving the micro-motion platform 2 (202), so that a corresponding relationship can be generated according to the displacement of the micro-motion platform 2 (202) and the drift of the central wavelength of the optical fiber grating reflection spectrum.

4. The method for selecting a demodulation model for a fiber Bragg grating stress sensor based on calibration data according to claim 1, characterized in that: The internal parameters of the neural network are continuously debugged to achieve the optimal effect, including the following parameters: the number x and size y of convolution kernels, stride z; the type o (such as maximum pooling or average pooling) and window size v of the pooling layer (303); the number r of neurons in the fully connected layer (304); the type p of the activation function (such as ReLU, Sigmoid); and the learning rate j of the optimizer; by continuously optimizing the parameters, after demodulation, the following sequence is obtained, in, represents the estimated value of the central wavelength under the lth stress condition, and cnn represents the convolutional neural network model used.

5. The method for selecting a demodulation model for a fiber Bragg grating stress sensor based on calibration data according to claim 1, characterized in that: The output of the neural network is the root mean square error (RMSE) after training with different parameters. RMSE is used to measure the difference between the model prediction value and the actual value. Its formula is: By calculating the demodulation results W of (KA) fiber Bragg grating sensors (103) under different parameters cnn(L) , and the input learning sequence W l (M), the minimum RMSE value obtained by training KA sensors is recorded as {RMSE1, RMSE2, ..., RMSE K-A }, the B FBG sensors with the best results reaching RMSE < 0.01 are classified as the ones that can be trained with neural networks.

6. The method for selecting a demodulation model for a fiber Bragg grating stress sensor based on calibration data according to claim 1, characterized in that: The input layer (301) of the neural network is based on the spectral data under each stress condition (each spectrum has Sampling points), first read and standardize the fiber Bragg grating sensor (103) data, and then divide it into input (x) and output (y); the data set is divided into 90% training set and 10% test set, and managed by using a custom data set class and data loader; select a suitable computing device (CPU or GPU) for training, optimize the parameter calibration of the sensor through a neural network, and finally improve the accuracy of the fiber Bragg grating stress sensor and the reliability of engineering applications; The one-dimensional convolution is used in the neural network. Its operation is shown in the formula, where x(t) is the input signal, w(s) is the convolution kernel, and the convolution kernel size is assumed to be n. To extract features from input data through convolution operation, it is necessary to apply activation functions, such as ReLU, Sigmoid, Tanh, etc., to perform nonlinear transformation on the convolution output. The advantage of ReLU function is that it increases the sparsity of the network, has good generalization performance and small amount of calculation. The disadvantage is that sparse processing will lead to the inability to learn effective features, so we choose ReLU function. ReLU(x)=max(0,x) After the activation function, a pooling operation (such as maximum pooling or average pooling) may be performed to reduce the spatial size of the feature map and reduce the complexity of the model. Assuming that the pooling kernel size is 2, the expressions of one-dimensional maximum pooling and one-dimensional average pooling are as follows. We choose maximum pooling in the network because maximum pooling focuses more on retaining the significant part of the feature and ignoring other information. P(t)=max[y′(2t),y′(2t+1)] The fully connected layer (304) is the last layer of CNN, which is responsible for organizing the feature information after pooling, extracting and saving useful information; wherein P is the output of the previous layer, W is the weight of the fully connected layer (304), and B is the bias; the process of implementing the fully connected layer (304) embodies linear characteristics because it performs calculations by linearly combining the input and weights. Y=W.P+B 7. The method for selecting a demodulation model for a fiber Bragg grating stress sensor based on calibration data according to claim 1, characterized in that: With this experimental platform, users can easily replace and test different FBG sensors, whether they are produced in the same batch or in different packages. These FBG sensors will inevitably have slight differences in the production process, such as packaging methods, material properties or manufacturing processes. This feature enables the system to effectively deal with small differences between sensors and ensure accurate and consistent calibration of sensors with various packaging specifications. This flexibility enables users to select and apply the most suitable sensors in different stress measurement and monitoring scenarios, thereby improving the applicability and practicality of the system.

8. The method for selecting a demodulation model for a fiber Bragg grating stress sensor based on calibration data according to claim 1, characterized in that: In the database, key physical parameters such as stress value, wavelength movement, fit and root mean square error are recorded and stored in the database; in addition, the optimized neural network model parameters are also saved for direct call and application in subsequent engineering applications; users can use the query function of the database to instantly obtain the required parameter data according to the sensor identification or specific test conditions, so as to use this information in engineering applications and scientific research experiments to optimize sensor performance and improve its application effect and accuracy.

Images