Low-temperature structure stress compensation method based on neural network

Abstract

The invention discloses a low-temperature structural stress compensation method based on a neural network, and the method comprises the steps: collecting the wavelength data of an optical fiber strain sensor and the wavelength data of an optical fiber temperature sensor under different temperature and load conditions; calculating the wavelength variation of the optical fiber strain sensor and the wavelength variation of the optical fiber temperature sensor; constructing a BP neural network model of which input layer neurons are the wavelength variation of the optical fiber strain sensor and the wavelength variation of the optical fiber temperature sensor; setting training parameters of the BP neural network; and training and verifying the BP neural network model. By expanding the data range, introducing the data processing technology and the neural network training technology and changing the parameters of the input layer of the neural network, a more accurate, reliable and universal compensation method is provided for structural stress monitoring in the low-temperature environment, and the method has wide application prospects and practical application value.

Description

Technical Field

[0001] The invention relates to the technical field of data error compensation processing, and in particular to a low-temperature structure stress compensation method based on a neural network. Background Art

[0002] Artificial neural network is an abstract model that simulates the human brain in structure and function. It is composed of a large number of simple processing units - neurons, and has the characteristics of nonlinearity, self-learning ability, and adaptive distributed information processing. After more than half a century of twists and turns, the research on artificial neural networks has made significant progress and breakthroughs, and has attracted more and more attention. It has been widely used in pattern recognition, automatic control, signal processing, sensor technology, system modeling and other fields. The basic unit of artificial neural network is the neuron. The general model structure is as follows: Figure 1 shown.

[0003] Neural networks are generally composed of three parts: input layer, hidden layer and output layer. The input layer receives external signals, the hidden layer processes the signals and extracts features, and the output layer generates the final output signal. The input layer and the output layer are both one-layer structures, and the number of layers in the hidden layer structure is adjusted according to the complexity of the problem to be solved, which can be one or more layers.

[0004] There are many neural network models. According to the network structure, neural networks are divided into feedforward networks without feedback and feedforward networks. For feedforward networks, also known as feedforward networks, signals are input from the input layer and transmitted to the output layer through the hidden layer. Each layer of neurons only receives input from the previous layer of neurons and transmits the signal to the next layer. The next layer of neurons has no effect on the previous layer of neurons. The general structure of the feedforward network is as follows: Figure 2 As shown in (a) in the figure. For a feedback network, the information transmission relationship between neurons is not only a single layer-by-layer transmission, but also the latter layer has a feedback effect on the previous layer. The structure is more complex than the forward network. The general structure of the feedback network is as follows: Figure 2 As shown in (b) in .

[0005] BP (Back-Propagation) neural network is the most widely used ANN model at present. It belongs to the forward network and has strong nonlinear mapping ability. The training rule of BP neural network is: during the training process, according to the steepest descent method, the error value between the actual output and the expected output of the training network is compared, and then the weights and thresholds of the network are optimized layer by layer along the negative gradient direction of the error change to achieve the minimum sum of squared errors.

[0006] The specific working process of BP neural network is as follows: there will be a certain difference between the actual output value of the network and the expected output value. The neural network reduces this difference by adjusting the connection weights to complete the training. The specific training algorithm is: Assume that the input of the network is x1(p), x2(p),..., x n (p), the actual output is y(p), and the expected output is y d (p), by taking the input x1(p),x2(p),...,x n (p) and the expected output y d (p) The neuron is activated and begins iterative training. The actual output of the iterative calculation is: Where function is the activation function. Then the training error e(p) is: e(p) = y d (p)-y(p), if e(p) is a positive value, that is, the expected output is greater than the actual output, then the weight is adjusted to increase the actual output y(p); if e(p) is a negative value, that is, the expected output is less than the actual output, then the weight is adjusted to reduce the actual output y(p). The weight adjustment rule is: i (p+1)=w i (p)+Δw i (p), where Δw i (p) is the correction weight of iterative p, and the weight correction is calculated by gradient descent method: Δw i (p) = α × x i (p)×e(p). Repeat the above iterative process until the neural network reaches convergence.

[0007] The strain compensation method based on BP neural network in the existing technology mainly compensates for the structural strain in the temperature range of -18 to 60°C. In this method, the number of neurons in the input layer is 2, the number of neurons in the hidden layer is 5, and the number of neurons in the output layer is 1. The strain value of the optical fiber strain sensor measured in the experiment and the current temperature are used as input samples, and the strain value measured by the resistive strain gauge at the same position in each working condition at normal temperature is used as the output sample, and the BP neural network is trained for each type of optical fiber strain sensor. Before training, several samples are randomly selected as test samples to test the generalization ability of the optical fiber strain sensor, and the remaining samples are used as training samples for neural network training. The strain values ​​of some test samples of optical fiber strain sensors before and after compensation are shown in the figure below. Figure 3 shown.

[0008] The error of the optical fiber strain sensor after compensation is significantly reduced, and the average absolute error is reduced from 93.8 microstrain to 32.4 microstrain, which is 34.5% of the original average error. After BP neural network compensation, the error of the optical fiber strain sensor is significantly improved.

[0009] Although the existing BP neural network compensation algorithm can reduce the error of optical fiber strain sensor to a certain extent, it still has some limitations when it comes to compensating for low-temperature structural stress:

[0010] 1. Missing low-temperature data: The test data used in the training of existing neural network models is in the range of -18 to 60°C. The low-temperature data below -18°C is missing. The error characteristics of optical fiber strain sensors at low temperatures are different. For structural stress compensation in low-temperature environments, the existing neural network method may not be applicable.

[0011] 2. Insufficient generalization ability of neural networks: Existing neural network methods will have overfitting problems during training due to insufficient test data, resulting in insufficient generalization ability of the trained model and inability to be universally applied to strain compensation under various stress conditions.

[0012] 3. Inaccurate parameters of the neural network input layer: The input layer parameters of the existing neural network method are the strain value and temperature of the optical fiber strain sensor. The strain value of the optical fiber strain sensor is obtained by measuring the wavelength according to the wavelength-strain conversion formula given by the manufacturer. Due to the different installation methods and test component materials of the optical fiber strain sensor, the correlation coefficients in the wavelength-strain conversion formula will also be different, resulting in the converted strain value may not be accurate. Inaccurate input parameters will cause certain deviations in the training results of the neural network. Summary of the invention

[0013] The purpose of the present invention is to provide a method that can accurately compensate for structural stress measurement errors in a low-temperature environment. The method is based on the test measurement data of the optical fiber strain sensor obtained at low temperatures, and in view of the characteristics of the structural stress error, four data processing and neural network training technologies are added on the basis of the BP neural network, so as to enhance the universality of the compensation method, further improve the strain measurement accuracy of the optical fiber strain sensor under low-temperature conditions, solve the limitations of the existing technology in structural stress monitoring under low-temperature environments, and thus ensure the accuracy and reliability of ship structural stress monitoring.

[0014] The embodiment of the present invention provides a method for compensating stress of a cryogenic structure based on a neural network, comprising the following steps:

[0015] Step 1: Collect the wavelength data of the optical fiber strain sensor and the wavelength data of the optical fiber temperature sensor under different temperature and load conditions;

[0016] Step 2: Calculate the wavelength variation of the optical fiber strain sensor and the wavelength variation of the optical fiber temperature sensor;

[0017] Step 3: construct a BP neural network model whose input layer neurons are the wavelength variation of the optical fiber strain sensor and the wavelength variation of the optical fiber temperature sensor;

[0018] Step 4: Setting the training parameters of the BP neural network;

[0019] Step 5: training the BP neural network model;

[0020] Step 6: Verify the BP neural network model.

[0021] Furthermore, the step 1 of collecting the wavelength data of the optical fiber strain sensor and the wavelength data of the optical fiber temperature sensor includes: using a high and low temperature humidity test chamber to adjust the ambient temperature of the test piece, observing the surface temperature of the test piece measured by the thermocouple, and when the surface temperature of the test piece is stabilized at a desired temperature point, using a test machine to apply a load to the test piece starting from 0 according to a required load gradient, and the maximum load does not exceed the yield load of the test component;

[0022] The plateau data of each gradient load-keeping period are extracted, including the wavelength of the optical fiber strain sensor, the wavelength of the optical fiber temperature sensor, the strain value of the resistance strain gauge, the load size and the thermocouple temperature.

[0023] Furthermore, in step 2, the wavelength change of the optical fiber strain sensor = the real-time wavelength of the optical fiber strain sensor - the initial wavelength; the wavelength change of the optical fiber temperature sensor = the real-time wavelength of the optical fiber temperature sensor - the initial wavelength.

[0024] Furthermore, the construction of the BP neural network model in step 3 specifically includes: introducing L2 regularization technology, adding a penalty term in the loss function to limit the size of the model weight; the loss function of the L2 regularization is expressed as:

[0025]

[0026] Among them, MSE is the original loss function, w i is each weight parameter of the model, and λ is a hyperparameter that controls the strength of regularization. The larger the value, the stronger the effect.

[0027] Furthermore, constructing the BP neural network model in step 3 specifically includes: using the Dropout technique to randomly "discard" a portion of neurons during each training process, so that the model uses a different subset of neurons for training in each iteration;

[0028] The hyperparameter p is used to represent the dropout rate. During training, for each neuron in the hidden layer, the probability p determines whether to set its output to 0.

[0029] Among them, the output of the neurons that are not discarded is multiplied by 1 / (1-p) to keep the expected value of the output unchanged.

[0030] Furthermore, constructing the BP neural network model in step 3 specifically includes: during the training process of the model, monitoring the performance of the validation set is maintained, and when the validation set does not show any improvement within a set epoch, the training is terminated in advance.

[0031] Furthermore, the construction of the BP neural network model in step 3 specifically includes batch normalization technology: each batch of data is standardized so that they have similar distributions, and the BN layer is placed between the linear transformation and the nonlinear activation function;

[0032] First, calculate the mean μ and variance σ of each batch 2 :

[0033]

[0034] Where m is the size of each batch;

[0035] The input data is then normalized using the batch mean and variance to have a mean of 0 and a variance of 1:

[0036]

[0037] Among them, ∈ is a small constant used to prevent the denominator from being zero;

[0038] At the same time, learnable parameters γ and β for scaling and translation are added to prevent batch normalization from limiting the expressiveness of the model, so that the normalized output is rescaled to an arbitrary mean and variance:

[0039] y i =γ·x i +β

[0040] By learning the parameters γ and β, the model maintains the normalized results or restores to the original distribution;

[0041] Finally, the batch normalized output y i Pass it to the next layer.

[0042] Furthermore, setting the training parameters of the BP neural network in step 4 includes: determining the training parameters of the BP neural network, including a learning rate and a maximum number of training times.

[0043] Furthermore, the training of the BP neural network model in step 5 includes: using a portion of the preprocessed data to train the BP neural network, and optimizing the network weights through iteration so that the output strain value is close to the actual strain measurement reference value.

[0044] Furthermore, verifying the BP neural network model in step 6 includes: using another part of data that has not participated in the training to verify the trained BP neural network model, and evaluating the generalization ability and compensation effect of the model.

[0045] The beneficial effects of the present invention are:

[0046] 1. Expand the temperature application range of the strain compensation method: Add structural strain data in the range of -55 to -20°C, and expand the temperature application range of the neural network strain compensation method to adapt to a wider range of application scenarios.

[0047] 2. Enhance the generalization ability of neural networks: Through the four data processing and neural network training technologies of L2 regularization, Dropout, early stopping and batch normalization, the overfitting problem of the trained model can be alleviated and the generalization ability of the neural network model can be enhanced.

[0048] 3. Improve the accuracy of the neural network input layer parameters: In order to prevent the strain value of the optical fiber strain sensor from being affected by various external factors and to ensure the unique correspondence between the input layer parameters and the output layer parameters, the wavelength change (real-time wavelength - initial wavelength) of the optical fiber strain sensor and the optical fiber temperature sensor is used as the neural network input layer parameter to improve the accuracy of the neural network model training results and further improve the accuracy of the strain compensation method.

[0049] In summary, the present invention provides a more accurate, reliable and universal compensation method for structural stress monitoring in low-temperature environments by expanding the data range, introducing four data processing technologies and neural network training technologies, and changing the parameters of the neural network input layer. It has broad application prospects and practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A neuron model for the prior art;

[0051] Figure 2 A neural network model for the prior art;

[0052] Figure 3 Compensating fore and aft responses of a portion of a test specimen of a prior art strain sensor;

[0053] Figure 4 It is the structural stress compensation model of the present invention;

[0054] Figure 5 The flowchart of the low temperature structure stress compensation based on neural network of the present invention;

[0055] Figure 6 are the strain values ​​before and after compensation and the reference resistor strain value of the optical fiber strain sensor of the present invention;

[0056] Figure 7 This is the result of the low-temperature structural stress compensation based on the neural network of the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0058] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0059] It can be understood that the principle of BP neural network compensation for optical fiber strain sensor is to connect a BP neural network behind the response output of the optical fiber strain sensor to eliminate the influence of temperature on the optical fiber strain sensor. Assuming the ambient temperature is t, the input value of the optical fiber strain sensor is x (strain), the response value of the sensor is y (wavelength), and the output value x' (strain) of the BP neural network should be close to the measured value x in an ideal state. The structural stress compensation model is as follows Figure 4 As shown, Figure 4 In the figure: x is the target strain parameter to be measured; t is the non-target temperature parameter; y is the output wavelength of the optical fiber strain sensor; x' is the strain value after BP neural network compensation; considering that the initial wavelengths of each optical fiber strain sensor are different after installation, the wavelength change of the optical fiber strain sensor (real-time wavelength-initial wavelength) is used as the input parameter y, and the wavelength change of the optical fiber temperature sensor (real-time wavelength-initial wavelength) is used as the input parameter t.

[0060] The technical solution of the present invention mainly includes the following steps:

[0061] Step 1. Dataset generation:

[0062] Conduct temperature-strain coupling test, the temperature range covers the required temperature range of the application environment, and the temperature gradient is 10℃; use high and low temperature humidity test chamber to adjust the ambient temperature of the test piece, observe the surface temperature of the test piece measured by the thermocouple, wait for the surface temperature of the test piece to stabilize at the required temperature point, and the fluctuation does not exceed 1℃. After 5 minutes, use the test machine to apply load to the test piece from 0 according to the required load gradient. The maximum load does not exceed the yield load of the test component. After completing a set of loading, unload the test machine, adjust the temperature, and continue to apply load from 0 after the temperature of the test chamber stabilizes. Collect wavelength data of optical fiber strain sensors under different temperature and load conditions.

[0063] Step 2. Data preprocessing:

[0064] At each temperature, each load is applied continuously from 0 to the maximum load in a gradient manner. The holding time of each load gradient is 10s. The stable segment data of each gradient holding time is extracted, including the wavelength of the optical fiber strain sensor, the wavelength of the optical fiber temperature sensor, the strain value of the resistance strain gauge, the load size and the thermocouple temperature; the wavelength change of the optical fiber strain sensor (real-time wavelength-initial wavelength) and the wavelength change of the optical fiber temperature sensor (real-time wavelength-initial wavelength) are calculated.

[0065] Step 3. Neural network model construction:

[0066] A three-layer BP neural network is constructed, with 2 neurons in the input layer (optical fiber strain sensor wavelength change and optical fiber temperature sensor wavelength change), 1 neuron in the output layer (compensated strain value), and 5 neurons in the hidden layer. The above basic BP neural network can preliminarily simulate the nonlinear relationship between input and output parameters, but considering the test time and economic cost, it is not possible to obtain enough data, resulting in insufficient generalization ability of the trained model and failure to obtain good results on the test set. Therefore, the following four data processing and neural network training technologies are introduced:

[0067] 1) L2 Regularization Technique

[0068] The goal of a neural network is to minimize the loss function. This neural network model uses mean square error (MSE). Without regularization, the model parameters (i.e., weights) can be adjusted freely to fit the training data as accurately as possible. However, this may overfit the training data and perform poorly on unseen test data. The L2 regularization technique is introduced to limit the size of the model weights by adding a penalty term to the loss function to reduce the risk of overfitting. The loss function of L2 regularization can be expressed as:

[0069]

[0070] Among them, MSE is the original loss function, w i is each weight parameter of the model, and λ is a hyperparameter that controls the strength of regularization. The larger the value, the stronger the effect. This penalty term makes the model tend to reduce the size of the weight during training, encouraging the weight to be close to zero, thereby reducing the complexity of the model and making it no longer overly dependent on certain specific features. In this way, the model is not prone to overfitting to noise or accidental features in the training data, and can perform more robustly on unseen data, thereby improving generalization ability.

[0071] 2) Dropout Technology

[0072] Dropout technology randomly "discards" a part of neurons during each training process, so that the model uses a different subset of neurons for training in each iteration. The hyperparameter p is used to represent the dropout rate. During training, for each neuron in the hidden layer, the probability p determines whether to set its output to 0. For the output of neurons that are not discarded, multiply it by 1 / (1-p) to keep the expected value of the output unchanged. Through repeated experiments, the p value that can achieve the best effect on the test set is determined.

[0073] This technology reduces the co-adaptability between neurons by randomly discarding neurons, so that the model can learn more general features. This allows the network to perform well without certain specific neurons and no longer rely on specific neurons or specific propagation paths, thereby preventing overfitting and improving generalization capabilities.

[0074] 3) Early Stopping Technique

[0075] During the model training process, the performance of the validation set is monitored. When the validation set does not improve within the set epoch, the training will be terminated early. This can effectively prevent the model from overfitting on the training set.

[0076] 4) Batch Normalization Technology

[0077] In order to deal with the gradient vanishing or gradient exploding problems that may exist during the training process, each batch of data is standardized to have a similar distribution, thereby reducing the data distribution changes between different layers in the model training, so as to achieve the purpose of accelerating training and improving the generalization ability of the model. The BN layer is placed between the linear transformation (hidden fully connected layer) and the nonlinear activation function (ReLU).

[0078] First, calculate the mean μ and variance σ of each batch 2 :

[0079]

[0080] Here, m is the size of each batch.

[0081] The input data is then normalized using the batch mean and variance to have a mean of 0 and a variance of 1:

[0082]

[0083] Here, ∈ is a small constant used to prevent the denominator from being zero.

[0084] At the same time, we can add learnable parameters γ and β for scaling and translation to prevent batch normalization from limiting the expressiveness of the model, so that the normalized output can be rescaled to any mean and variance:

[0085] y i =γ·x i +β

[0086] By learning γ and β, the model can choose to keep the normalized results or restore to the original distribution.

[0087] Finally, the batch normalized output y i Pass it to the next layer.

[0088] Step 4. Training parameter settings:

[0089] Determine the training parameters of the BP neural network, such as learning rate, maximum number of training times, etc. These parameters have an important impact on the convergence speed and final performance of the model.

[0090] Step 5. Neural network model training:

[0091] A portion of the preprocessed data is used to train the BP neural network. The network weights are optimized iteratively to make the output strain value as close as possible to the actual strain measurement reference value (resistance strain gauge strain value).

[0092] Step 6. Neural network model verification:

[0093] Use another part of the data that was not involved in the training to verify the trained model and evaluate the generalization ability and compensation effect of the model.

[0094] Step 7. Application of neural network compensation method:

[0095] The trained BP neural network is applied to the actual low-temperature structure stress monitoring to compensate the measurement error of the optical fiber strain sensor and provide accurate stress monitoring data.

[0096] Figure 5 The figure is a flow chart of stress compensation for cryogenic structures based on neural networks. The figure describes in detail the entire processing process from sensor data input to strain value output after compensation. In the figure: x is the target structure strain to be measured; t is the output wavelength of the temperature sensor; y is the output wavelength of the optical fiber strain sensor; x' is the strain value after BP neural network compensation; considering that the initial wavelengths of each optical fiber sensor after installation are somewhat different, the measured data is preprocessed, and the wavelength change of the optical fiber strain sensor (real-time wavelength-initial wavelength) is used as the input parameter y' of the neural network compensation model, and the wavelength change of the optical fiber temperature sensor (real-time wavelength-initial wavelength) is used as the input parameter t'.

[0097] Figure 6The strain values ​​of the optical fiber strain sensor before and after compensation by the neural network method and the reference resistance strain gauge strain value are shown in the range of -55℃ to 50℃ under three loading conditions (30, 60, and 90kN). The difference between the strain value of the optical fiber strain sensor after compensation and the strain value of the resistance strain gauge is significantly reduced compared with that before compensation.

[0098] Example

[0099] In order to fully disclose the technical content of the present invention, a specific embodiment of a low-temperature structure stress compensation method based on a neural network is provided below:

[0100] 1. Dataset generation: Temperature-strain coupling tests were carried out, and tensile or bending tests were performed on the test pieces at twenty-two temperature points, namely -55℃, -50℃, -40℃, -30℃, -20℃, -10℃, 0℃, 10℃, 20℃, 30℃, 40℃, 50℃, 40℃, 30℃, 20℃, 10℃, 0℃, -10℃, -20℃, -30℃, -40℃, and -50℃ (the first twelve temperature points are the heating stage, and the last ten temperature points are the cooling stage). Use a high and low temperature humidity test chamber to adjust the ambient temperature of the test piece, observe the surface temperature of the test piece measured by the thermocouple, wait for the surface temperature of the test piece to stabilize at the required temperature point, and fluctuate no more than 1°C for 5 minutes, then use the test machine to apply load to the test piece from 0 according to the required load gradient, and the maximum load shall not exceed the yield load of the test component. After completing a set of loading, unload the test machine, adjust the temperature, and continue to apply load from 0 after the temperature of the test chamber stabilizes. Collect the wavelength data of the optical fiber strain sensor under different temperature and load conditions. See Table 1 below for detailed test information:

[0101] Table 1 Test load information

[0102]

[0103]

[0104] 2. Data preprocessing: At each temperature, each load is applied continuously from 0 to the maximum load in a gradient, and the holding time of each load gradient is 10s. The stable segment data of each gradient holding is extracted, including the wavelength of the optical fiber strain sensor, the wavelength of the optical fiber temperature sensor, the strain value of the resistance strain gauge, the load size and the thermocouple temperature. Taking the test piece LSSYJ01 at -30℃ as an example, the data record is shown in Table 2. The above test is carried out on each test piece in Table 1 and the data is recorded; the change in the wavelength of the optical fiber strain sensor and the change in the wavelength of the optical fiber temperature sensor (real-time wavelength-initial wavelength) are calculated.

[0105] Table 2 Test record

[0106]

[0107]

[0108]

[0109] 3. Neural network model construction: A three-layer BP neural network is constructed, with 2 neurons in the input layer (wavelength change of optical fiber strain sensor and wavelength change of optical fiber temperature sensor), 1 neuron in the output layer (strain value after compensation), 1 hidden layer, and 5 neurons. The probability p in the Dropout technique is set to 0.5, and the training is terminated early when there is no improvement within 200 epochs in the early stopping technique.

[0110] 4. Training parameter setting: Determine the training parameters of the BP neural network, with a learning rate of 0.001 and a maximum training number of 20,000 epochs.

[0111] 5. Neural network model training: Use a portion of the preprocessed data to train the BP neural network. By iteratively optimizing the network weights, the output strain value is made as close as possible to the actual strain measurement reference value (resistance strain gauge strain value).

[0112] 6. Neural network model verification: Use another part of the data that has not participated in the training to verify the trained model and evaluate the generalization ability and compensation effect of the model. The specific compensation effects are shown in Table 3 and Figure 7 shown.

[0113] Table 3 Results of low temperature structure stress compensation based on neural network

[0114]

[0115]

[0116]

[0117]

[0118] For a certain type of optical fiber strain sensor, a low-temperature stress compensation method based on neural network is used to compensate the measured values ​​of the optical fiber strain sensor at different temperatures and different loads. From the compensation results, it can be seen that this method has good compensation accuracy for structural stress at low temperatures, and the strain error after compensation is greatly reduced. In the range of -55℃~50℃, the mean values ​​of the difference after compensation under three loading conditions are 9 microstrain, 19 microstrain, and 25 microstrain, respectively, and the maximum error reductions are 254.92%, 126.22%, and 84.47%, respectively, which proves the practical value of this method.

[0119] The present invention achieves significant technical effects of structural stress compensation at low temperatures by adopting the BP neural network method. The following are the main technical effects of the present invention:

[0120] 1. The temperature application range of the structural stress compensation method is wide: within the temperature range of -55 to 50°C, the structural stress compensation method based on the neural network is applicable.

[0121] 2. The structural stress compensation method has strong universality: Through the four data processing and neural network training technologies of L2 regularization, Dropout, early stopping and batch normalization, the overfitting problem of the trained model is alleviated, the generalization ability of the neural network model is greatly enhanced, and the universality of the structural stress compensation method is further improved.

[0122] 3. High accuracy of structural stress compensation results: By adjusting the neural network input parameters and introducing four data processing and neural network training technologies, the accuracy of the neural network model training results is improved, and the accuracy of the strain compensation method is further improved.

[0123] Through the description of the above technical effects, it can be seen that the present invention is not only innovative in theory, but also has significant practical value in practical applications. These effects not only improve the accuracy and reliability of stress monitoring, but also provide new possibilities for technological progress and application in related fields.

[0124] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for compensating low temperature structure stress based on neural network, characterized in that: The steps include: Step 1: Collect the wavelength data of the optical fiber strain sensor and the wavelength data of the optical fiber temperature sensor under different temperature and load conditions; Step 2: Calculate the wavelength variation of the optical fiber strain sensor and the wavelength variation of the optical fiber temperature sensor; Step 3: construct a BP neural network model whose input layer neurons are the wavelength variation of the optical fiber strain sensor and the wavelength variation of the optical fiber temperature sensor; Step 4: Setting the training parameters of the BP neural network; Step 5: training the BP neural network model; Step 6: Verify the BP neural network model.

2. The method according to claim 1, characterized in that Collecting the wavelength data of the optical fiber strain sensor and the wavelength data of the optical fiber temperature sensor in step 1 includes: using a high and low temperature humidity test chamber to adjust the ambient temperature of the test piece, observing the surface temperature of the test piece measured by the thermocouple, and when the surface temperature of the test piece is stabilized at a desired temperature point, using a test machine to apply a load to the test piece starting from 0 according to a required load gradient, and the maximum load does not exceed the yield load of the test component; The plateau data of each gradient load-keeping period are extracted, including the wavelength of the optical fiber strain sensor, the wavelength of the optical fiber temperature sensor, the strain value of the resistance strain gauge, the load size and the thermocouple temperature.

3. The method according to claim 1, characterized in that In step 2, the wavelength change of the optical fiber strain sensor = the real-time wavelength of the optical fiber strain sensor - the initial wavelength; the wavelength change of the optical fiber temperature sensor = the real-time wavelength of the optical fiber temperature sensor - the initial wavelength.

4. The method according to claim 1, characterized in that The step 3 of constructing the BP neural network model specifically includes: introducing L2 regularization technology, adding a penalty term in the loss function to limit the size of the model weight; the loss function of the L2 regularization is expressed as: Among them, MSE is the original loss function, w i is each weight parameter of the model, and λ is a hyperparameter that controls the strength of regularization. The larger the value, the stronger the effect.

5. The method according to claim 1, characterized in that The step 3 of constructing the BP neural network model specifically includes: using the Dropout technique to randomly "discard" a portion of neurons during each training process, so that the model uses a different subset of neurons for training in each iteration; The hyperparameter p is used to represent the dropout rate. During training, for each neuron in the hidden layer, the probability p determines whether to set its output to 0. Among them, the output of the neurons that are not discarded is multiplied by 1 / (1-p) to keep the expected value of the output unchanged.

6. The method according to claim 1, characterized in that The step 3 of constructing the BP neural network model specifically includes: during the training process of the model, the performance of the validation set is monitored, and when the validation set does not show any improvement within the set epoch, the training is terminated in advance.

7. The method according to claim 1, characterized in that The BP neural network model constructed in step 3 specifically includes batch normalization technology: each batch of data is standardized so that they have similar distribution, and the BN layer is placed between the linear transformation and the nonlinear activation function; First, calculate the mean μ and variance σ of each batch 2 : Where m is the size of each batch; The input data is then normalized using the batch mean and variance to have a mean of 0 and a variance of 1: Among them, ∈ is a small constant used to prevent the denominator from being zero; At the same time, learnable parameters γ and β for scaling and translation are added to prevent batch normalization from limiting the expressiveness of the model, so that the normalized output is rescaled to an arbitrary mean and variance: y i =γ·x i +b By learning the parameters γ and β, the model maintains the normalized results or restores to the original distribution; Finally, the batch normalized output y i Pass it to the next layer.

8. The method according to claim 1, characterized in that Setting the training parameters of the BP neural network in step 4 includes: determining the training parameters of the BP neural network, including a learning rate and a maximum number of training times.

9. The method according to claim 1, characterized in that: The training of the BP neural network model in step 5 includes: using a portion of the preprocessed data to train the BP neural network, and optimizing the network weights through iteration so that the output strain value is close to the actual strain measurement reference value.

10. The method according to claim 9, characterized in that Verifying the BP neural network model in step 6 includes: using another part of data that has not participated in the training to verify the trained BP neural network model, and evaluating the generalization ability and compensation effect of the model.

Images