Multi-core fiber grating multi-parameter decoupling method based on improved CNN-LSTM

Abstract

The invention relates to the technical field of optical fiber performance analysis, in particular to a multi-core fiber grating multi-parameter decoupling method based on improved CNN-LSTM. According to the method, a multi-core fiber grating temperature, vibration and three-dimensional deformation detection system is built, center wavelength data under the action of different physical loads are collected, and the center wavelength data are matched with corresponding real physical quantities to build a supervised learning data set. After abnormal value elimination, filtering noise reduction and normalization processing are carried out on collected signals, stratified sampling and data enhancement are adopted to construct a training set and a test set. An improved CNN-LSTM model is established, multi-parameter nonlinear mapping and feature focusing are realized in combination with convolutional feature extraction, time sequence modeling and an attention mechanism, training is performed by taking RMSE as a loss function and an Adam optimizer, and an early stop mechanism is set to prevent overfitting. After training is completed, wavelength data to be measured are input into the model, and temperature, vibration and three-dimensional deformation results after decoupling can be output.

Description

Technical Field

[0001] This invention relates to the field of optical fiber performance analysis technology, and in particular to a multi-parameter decoupling method for multi-core fiber gratings based on an improved CNN-LSTM. Background Technology

[0002] Traditional fiber optic grating sensing technology primarily uses a single-core optical fiber as the interference medium, sensing physical quantities such as temperature, strain, vibration, pressure, displacement, and acceleration by measuring Bragg wavelength shift. This technology boasts advantages such as small size, light weight, resistance to electromagnetic interference, high sensitivity, high compatibility, and corrosion resistance, demonstrating significant advantages in the field of optical sensing. However, in the context of modern scientific research and engineering practice, the measurement of a single physical quantity is insufficient to meet the demands for in-depth understanding, precise control, and safety assurance of complex systems. Single-core structures struggle to simultaneously measure multiple physical quantities in complex environments. Researchers often integrate multiple sensing technologies into a single fiber core to achieve the measurement of multiple physical quantities; however, challenges arise in signal crosstalk and multi-parameter decoupling, increasing the measurement difficulty. Furthermore, practical measurements typically require various different optical fibers and supporting equipment, complicating the sensing system structure. Currently, the demand for multi-parameter sensing in fiber optic sensing technology is increasing, becoming a crucial trend in the modern sensing field.

[0003] In recent years, with the advancement of optical fiber manufacturing technology, multi-core optical fibers have attracted widespread attention due to their unique spatial structure advantages. Multi-core optical fibers integrate multiple cores within a single fiber, each of which can independently function as a sensing unit or carry different sensing mechanisms. This unique structure enables the simultaneous acquisition of multiple physical quantities such as temperature, strain, vibration, and curvature at a single measurement point, improving sensing efficiency and information density.

[0004] Currently, traditional multi-parameter decoupling methods include dual-grating methods, sensitivity matrices, multinomial regression, reference grating differential, MZ interferometry, and edge filtering. These methods generally employ a "segmented processing" strategy: first finding peaks, then filtering, then fitting or matrix inversion. Errors at each stage are propagated and accumulated, leading to a decrease in the final decoupling accuracy. This invention uses an improved CNN-LSTM network to achieve end-to-end mapping output, compressing the entire decoupling process into a single forward propagation. This eliminates the need for manual segmentation, significantly reducing accumulated errors. Simultaneously, it combines an attention mechanism to precisely focus key information in the LSTM network and automatically assign dynamic weights, achieving rapid and accurate decoupling of temperature, vibration, and three-dimensional deformation data. Summary of the Invention

[0005] To overcome the above shortcomings, this invention provides a multi-parameter decoupling method for multi-core fiber optic gratings based on an improved CNN-LSTM, aiming to address the problems of signal coupling, low decoupling accuracy, and insufficient model generalization ability in existing technologies when multi-core fiber optic gratings are simultaneously measured with multiple parameters.

[0006] This invention provides the following technical solution: a multi-parameter decoupling method for multi-core fiber gratings based on an improved CNN-LSTM, comprising:

[0007] Step 1: Build a multi-core fiber grating temperature, vibration, and three-dimensional deformation detection system, and acquire the center wavelength data when subjected to temperature, vibration, and three-dimensional deformation respectively, and generate a supervised learning dataset;

[0008] Step 2: Preprocess the collected center wavelength data of the multi-core fiber grating, and use the MinMax normalization method to normalize the center wavelength data of each fiber core to the [0,1] interval;

[0009] Step 3: Divide the normalized center wavelength data into training and test sets in a 7:3 ratio, and construct the training dataset by combining stratified sampling and data augmentation.

[0010] Step 4: Build and train the improved CNN-LSTM network model, and evaluate the test dataset using the trained CNN-LSTM model;

[0011] Step 5: Input the center wavelength data of the multi-core fiber grating to be tested into the trained CNN-LSTM model, and output the decoupled temperature, vibration and three-dimensional deformation data.

[0012] Preferably, the detection system consists of a multi-core fiber grating, a fan-in / fan-out module, a fiber grating demodulator, a heating device, a thermistor, a temperature measuring instrument, a vibration platform, an accelerometer, an acceleration measuring instrument, a standard bending mold, and a cantilever beam.

[0013] Preferably, the steps for obtaining the supervised learning dataset include:

[0014] A temperature load is applied to the cantilever beam by a heating device, and a temperature sensor collects temperature change data in real time as the true temperature value.

[0015] A vibration load is applied to the cantilever beam by a vibration excitation device, and an acceleration sensor collects the vibration acceleration data in real time as the true value of the vibration.

[0016] A three-dimensional deformation load with a preset curvature is applied to a multi-core fiber optic grating using a standard bending die, and the true value of the three-dimensional deformation is obtained based on the geometric parameters of the die.

[0017] The center wavelength data of each fiber core of the multi-core fiber grating is collected synchronously under various loads. The center wavelength data is then paired with the corresponding true values ​​of temperature, vibration, and three-dimensional deformation to construct a supervised learning dataset.

[0018] Preferably, the preprocessing step for the center wavelength data includes:

[0019] Outlier detection and removal are performed on the collected center wavelength data;

[0020] The center wavelength data is filtered to remove high-frequency noise and interference signals.

[0021] Preferably, the steps for constructing the training dataset include:

[0022] A stratified sampling method was adopted to stratify the data according to different working conditions such as temperature, vibration and deformation, ensuring that the data of each stratum were distributed proportionally to the training set and the test set.

[0023] Data augmentation is performed on the training set. Augmentation methods include time shifting, amplitude scaling, adding Gaussian noise, or time warping to increase the number of training samples.

[0024] Preferably, the CNN-LSTM network model includes an input layer, hidden layers, attention layers, fully connected layers, and an output layer, wherein:

[0025] The input layer receives the normalized center wavelength data;

[0026] The hidden layer includes a CNN convolutional layer and an LSTM network layer. The CNN convolutional layer is used to extract spatial features, and the LSTM network layer extracts temporal features through a gating mechanism.

[0027] The attention layer calculates attention weights for the hidden state sequence output by the LSTM network layer;

[0028] The fully connected layer integrates the weighted features;

[0029] The output layer outputs the decoupled temperature value, vibration value, and three-dimensional deformation value.

[0030] Preferably, the attention weight calculation process of the attention layer includes:

[0031] Construct a matching scoring function for the hidden state at each time step of the LSTM network layer output and the hidden state of the output layer;

[0032] The scoring results are normalized using the softmax function to obtain the attention weight vector.

[0033] Preferably, during model training, the root mean square error (RMSE) is used as the loss function; the Adam optimizer is used to update the network weights through backpropagation algorithm, and the learning rate is set to 0.001-0.01; an early stopping mechanism is set, and training is stopped when the validation set loss does not decrease for 5-10 consecutive rounds to prevent overfitting.

[0034] Preferably, the following evaluation metrics are used to evaluate model performance:

[0035] The coefficient of determination is used to evaluate the goodness of fit of a model to the data.

[0036] Root mean square error (RMSE) measures the degree of deviation between the predicted value and the actual value.

[0037] Average deviation is used to assess the systematic bias in predictions.

[0038] The present invention has the following beneficial effects:

[0039] 1. This invention, based on the CNN-LSTM deep learning architecture, fully leverages the advantages of convolutional networks in spatial feature extraction and long short-term memory networks in temporal feature modeling. It enables end-to-end nonlinear mapping analysis of complex signals generated by multi-core fiber gratings under conditions of temperature, vibration, and three-dimensional deformation coupling. The algorithm possesses strong self-learning and feature extraction capabilities, achieving automatic identification and separation of multi-parameter information without manual decoupling. This significantly improves the real-time performance and accuracy of fiber grating measurements in complex environments, making it suitable for scenarios such as high-dynamic structural health monitoring and multi-physics sensing.

[0040] 2. The improved CNN-LSTM decoupling method proposed in this invention adopts an end-to-end mapping mechanism, which effectively reduces the error accumulation between computational steps in traditional decoupling algorithms. By directly establishing a functional mapping relationship between the center wavelength change and the target physical quantity, the model can achieve simultaneous prediction of multiple parameters in one forward propagation, significantly improving decoupling efficiency and stability.

[0041] 3. This invention introduces an attention mechanism module into the LSTM network, enabling the model to dynamically allocate weights based on the importance of temporal features, thus differing from the equal-weighted memory method of traditional LSTM. This mechanism can automatically focus on the temporal features that contribute most to temperature, vibration, or three-dimensional deformation, enhancing the expressive power of key features and suppressing redundant information interference. Attached Figure Description

[0042] Figure 1 This is a flowchart of a multi-core fiber optic grating multi-parameter decoupling method based on an improved CNN-LSTM proposed in this invention;

[0043] Figure 2 This is a structural diagram of the detection system provided in an embodiment of the present invention;

[0044] Figure 3 The center wavelengths of the pre-processed middle core and side cores provided in the embodiments of the present invention;

[0045] Figure 4 A diagram illustrating the improved CNN-LSTM model structure provided in this embodiment of the invention;

[0046] Figure 5 The diagram shows the model structure of LSTM provided in this embodiment of the invention. Detailed Implementation

[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] This invention provides a multi-parameter decoupling method for multi-core fiber gratings based on an improved CNN-LSTM, such as... Figure 1 As shown, it includes the following steps:

[0049] Step 1: Build a multi-core fiber grating temperature, vibration, and three-dimensional deformation detection system, and acquire the center wavelength data when subjected to temperature, vibration, and three-dimensional deformation respectively, and generate a supervised learning dataset;

[0050] Preferably, the detection system consists of a multi-core fiber grating, a fan-in / fan-out module, a fiber grating demodulator, a heating device, a thermistor, a temperature measuring instrument, a vibration platform, an accelerometer, an acceleration measuring instrument, a standard bending mold, and a cantilever beam.

[0051] Specifically, a multi-core fiber grating temperature, vibration, and three-dimensional deformation detection system was constructed. The experimental system is as follows: Figure 2 As shown, the system includes a multi-core fiber Bragg grating, fan-in / fan-out modules, a fiber Bragg grating demodulator, a heating device, a thermistor, a temperature measuring instrument, a vibration platform, an accelerometer, a standard bending mold, and a cantilever beam. The fiber Bragg grating demodulator has an accuracy of [value missing] and a demodulation speed of [value missing]. In the multi-core fiber optic sensing system, the number of fiber cores is n, where n=1 represents the middle core. Each fiber core is etched with m gratings, with a grating length of 6mm. The acquisition time for each fiber core is T, the acquisition frame rate is K, and the input data is the wavelength data of the multi-core fiber Bragg grating affected by temperature, vibration, and three-dimensional deformation.

[0052] The heating device acts as a thermal excitation source, applying a specific temperature range to the cantilever beam to simulate the thermal environment the structure might encounter during actual operation, thereby studying the beam's temperature characteristics. The shaking table acts as a vibration excitation source, applying mechanical vibration to the cantilever beam to study its vibration modes, dynamic strain, etc. A standard bending die is equipped with multi-core optical fibers to acquire three-dimensional deformation data of controllable bending.

[0053] Preferably, the steps for obtaining the supervised learning dataset include:

[0054] A temperature load is applied to the cantilever beam by a heating device, and a temperature sensor collects temperature change data in real time as the true temperature value.

[0055] A vibration load is applied to the cantilever beam by a vibration excitation device, and an acceleration sensor collects the vibration acceleration data in real time as the true value of the vibration.

[0056] A three-dimensional deformation load with a preset curvature is applied to a multi-core fiber optic grating using a standard bending die, and the true value of the three-dimensional deformation is obtained based on the geometric parameters of the die.

[0057] The center wavelength data of each fiber core of the multi-core fiber grating is collected synchronously under various loads. The center wavelength data is then paired with the corresponding true values ​​of temperature, vibration, and three-dimensional deformation to construct a supervised learning dataset.

[0058] Specifically, in the CNN-LSTM network, temperature, vibration, and three-dimensional deformation data are acquired to serve as the supervised learning dataset for the CNN-LSTM network, thereby achieving controllability supervision of the network. More specifically, temperature, vibration, and three-dimensional deformation loads are applied to the fiber grating using experimental equipment such as heating devices, vibration platforms, and standard bending molds, and the corresponding ground truth values ​​are obtained, thus acquiring both training and supervised learning datasets based on the change in the center wavelength of the fiber grating.

[0059] Due to the unique sensing characteristics of multi-core fiber Bragg gratings, their side cores are sensitive to temperature, vibration, and curvature, while the middle cores are sensitive to temperature and vibration, but insensitive to curvature. Therefore, the expression for the center wavelength of the fiber Bragg grating output by the computer can be given as:

[0060]

[0061] For the middle core and The wavelength expression for the intermediate core is: For the side core n and The wavelength expression for the side core is: In the formula Indicates which fiber core, It is the middle core. It is the temperature coefficient. It is the strain coefficient. Indicates temperature. and These represent the strain caused by curvature and vibration, respectively. The center wavelength data is paired with the corresponding true values ​​of temperature, vibration, and three-dimensional deformation, i.e., these data are in the same time series, thus constructing a supervised learning dataset.

[0062] This step involves applying temperature, vibration, and three-dimensional deformation loads under controlled experimental conditions to acquire the response of multi-core fiber gratings under different physical actions, and accurately correlates the center wavelength change with the actual measured values ​​of each physical quantity to construct a high-quality supervised learning dataset.

[0063] Step 2: Preprocess the collected center wavelength data of the multi-core fiber grating, and use the MinMax normalization method to normalize the center wavelength data of each fiber core to the [0,1] interval;

[0064] Preferably, the preprocessing step for the center wavelength data includes:

[0065] Outlier detection and removal are performed on the collected center wavelength data;

[0066] The center wavelength data is filtered to remove high-frequency noise and interference signals.

[0067] Specifically, the acquired center wavelength data of multi-core fiber Bragg gratings may contain outliers caused by light source fluctuations, interference reflections, or sampling errors. To ensure data quality, statistical methods are first used to detect and remove outliers, including: calculating the mean value of the center wavelength time series for each fiber core. and standard deviation For any sampled value at any time If satisfied If it is, then it is determined to be an outlier. Typically, 2 or 3 is used; outliers are replaced by interpolation using the mean or median of adjacent non-outliers to ensure signal continuity.

[0068] To eliminate high-frequency noise and system interference in the center wavelength signal, the data is filtered. Specifically, a low-pass filter can be used; that is, for data with significant high-frequency interference, a first-order Butterworth low-pass filter can be used, with the cutoff frequency set to a value equal to the signal's dominant frequency. The final preprocessed center wavelength data is obtained by multiplying the previous values, as shown below. Figure 3 As shown.

[0069] After filtering, the center wavelength data is mapped to the [0,1] interval using the MinMax normalization method to eliminate dimensional differences under different fiber cores and measurement conditions, thereby improving the stability of model training. The normalization formula is as follows: ,in, This is the original center wavelength data. The result after normalization. and These are the maximum and minimum values ​​of the feature data, respectively.

[0070] The above preprocessing steps can effectively improve the quality and consistency of data, reduce the interference of outliers and noise signals on model training, and provide reliable input for the high-precision learning of CNN-LSTM models.

[0071] Step 3: Divide the normalized center wavelength data into training and test sets in a 7:3 ratio, and construct the training dataset by combining stratified sampling and data augmentation.

[0072] Preferably, the steps for constructing the training dataset include:

[0073] A stratified sampling method was adopted to stratify the data according to different working conditions such as temperature, vibration and deformation, ensuring that the data of each stratum were distributed proportionally to the training set and the test set.

[0074] Data augmentation is performed on the training set. Augmentation methods include time shifting, amplitude scaling, adding Gaussian noise, or time warping to increase the number of training samples.

[0075] Specifically, because the response amplitudes and distributions of temperature, vibration, and three-dimensional deformation differ under different operating conditions, directly randomly dividing the data might result in insufficient samples for some operating conditions in the training set, affecting the model's learning ability. To ensure a balanced data distribution, stratified sampling is used for sample division. The specific steps are as follows:

[0076] Based on the experimental conditions, the data was stratified according to the value ranges of three physical parameters. For example, the temperature range was divided into low, medium, and high temperatures based on temperature amplitude; the vibration range was divided into low, medium, and high frequencies based on acceleration amplitude; and the three-dimensional deformation was divided into small, medium, and large deformations based on curvature. Samples were randomly drawn from each range to maintain a 7:3 ratio between the training and test sets, while ensuring a consistent statistical distribution across all layers.

[0077] After obtaining the training set, to further improve the robustness and adaptability of the model, multi-dimensional data augmentation was performed on the training data. This included: shifting the center wavelength sequence by several sampling points in the time dimension to simulate changes in the sampling start point or signal phase shift, enhancing the model's tolerance to timing alignment errors; multiplying the center wavelength change by a random scaling factor to simulate differences in measurement system sensitivity, improving the model's adaptability to different measurement gains; and adding data with a mean of 0 and a standard deviation equal to the original signal mean to the center wavelength signal. Gaussian noise is used to simulate experimental noise and optical system interference, enhancing the model's noise resistance; slight nonlinear stretching or compression of time series data is applied to change the sampling interval without altering the overall trend, thereby improving the model's robustness to changes in sampling frequency.

[0078] The training dataset constructed through the above-mentioned stratified sampling and data augmentation can ensure a balanced distribution of samples under different working conditions, improve the stability of the model under disturbances such as noise and sensitivity bias, and significantly improve the generalization performance and prediction accuracy of the improved CNN-LSTM model.

[0079] Step 4: Build and train the improved CNN-LSTM network model, and evaluate the test dataset using the trained CNN-LSTM model;

[0080] Preferably, the CNN-LSTM network model includes an input layer, hidden layers, attention layers, fully connected layers, and an output layer, wherein:

[0081] The input layer receives the normalized center wavelength data;

[0082] The hidden layer includes a CNN convolutional layer and an LSTM network layer. The CNN convolutional layer is used to extract spatial features, and the LSTM network layer extracts temporal features through a gating mechanism.

[0083] The attention layer calculates attention weights for the hidden state sequence output by the LSTM network layer;

[0084] The fully connected layer integrates the weighted features;

[0085] The output layer outputs the decoupled temperature value, vibration value, and three-dimensional deformation value.

[0086] Preferably, the attention weight calculation process of the attention layer includes:

[0087] Construct a matching scoring function for the hidden state at each time step of the LSTM network layer output and the hidden state of the output layer;

[0088] The scoring results are normalized using the softmax function to obtain the attention weight vector.

[0089] Specifically, the model is as follows Figure 4 As shown, it can be divided into five main parts: the input layer is responsible for receiving and initializing the raw data; the hidden layer contains CNN and LSTM networks for in-depth analysis of the data and capturing temporal dependencies; the attention layer enhances the model's focus on key information by assigning different weights to different features; then there is the fully connected layer, which integrates the features extracted by the previous layers to support the final decoupling; finally, the output layer outputs the decoupled data of the model based on the processing results of the previous layers.

[0090] The input layer receives preprocessed center wavelength data containing temperature, vibration, and curvature information. To optimize the model initialization process, the Glorot algorithm is used to set the weights. This algorithm automatically adjusts the initial weight values ​​based on the input and output dimensions of the network layer. This helps maintain the stability of information flow in the network. The algorithm can be expressed as follows: In the formula, and These represent the number of input and output units for this layer, respectively.

[0091] Hidden layers primarily consist of CNN and LSTM networks. CNN convolutional layers extract spatial features by sliding multiple one-dimensional convolutional kernels across the input sequence. Through local connectivity and weight sharing mechanisms, spatial features are extracted from the data, and the kernels can be adjusted according to wavelength. The core of LSTM networks lies in the coordinated operation of cell states and three gating mechanisms. Information transmission is primarily regulated through the gating mechanism of the hidden layers. This gating mechanism consists of a forget gate, an input gate, and an output gate. Because its hidden layers contain four interaction layers that interact in unique ways, each gate has independent weights and thresholds. Figure 5 As shown, the detailed steps can be broken down as follows:

[0092] The core memory thread of the cellular state LSTM layer connects and runs through the entire time series, with only slight linear modifications at each time point, enabling information to be stably transmitted over long distances and is not easily lost.

[0093] The gating structure contains three gates, which are activated by three sigmoid functions and one tanh function. The output value is between 0 and 1 to regulate the information flow. A value of 0 represents "complete blocking", and a value of 1 represents "complete passage".

[0094] The forget gate determines the information that needs to be discarded from the previous cell state, and its calculation formula is as follows:

[0095]

[0096] in, The weight matrix representing the forget gate; This is the output of the previous time step; The input represents the current time step; It is the threshold matrix of the forget gate.

[0097] The input gate is the core of the LSTM, responsible for updating the cell with new information. It mainly involves two steps: first, the Sigmoid function determines the value to be updated; then, the Tanh function generates a candidate value vector, updates the cell layer, and creates a new cell state. The calculation formula can be expressed as:

[0098]

[0099]

[0100] In the formula, This represents the weight matrix of the input gate; It is the threshold matrix of the input gate; It is the weight matrix of the update unit; This represents the threshold matrix of the update unit.

[0101] After calculating the forget gate and input gate, the cell state update stage is required: the proportion of old information filtered out by the forget gate and the new candidate information allowed by the input gate are simultaneously integrated into the cell state, thereby incorporating the memory from the previous moment. Officially refreshed as the memory of the current moment. The specific update process is as follows:

[0102]

[0103] Finally, the output gate determines the output at the current moment, and the final output value can be represented as follows:

[0104]

[0105]

[0106] here, This represents the output gate weight matrix; This represents the output gate threshold matrix.

[0107] The attention layer receives the sequence of hidden states from the hidden layer at all time steps. It calculates a weight for each hidden state at each time step, reflecting the relevance and importance of the information at that time step to the current task. Based on these weights, the attention layer constructs a matching scoring function to evaluate the contribution of information from different time steps to the prediction.

[0108]

[0109] in, Indicates the first The hidden state output of each hidden layer This represents the hidden state output of the k-th output layer. It is a double tangent function, V, and This represents the attention model parameter matrix, which can be obtained through network training. At this point, the weight vectors of the fully connected layers of the CNN-LSTM network...

[0110]

[0111] here, It is a logistic regression function. Data extracted from the attention layer is integrated and mapped through a fully connected layer, and the output layer outputs relevant data on temperature, vibration, and curvature.

[0112] By implementing the above process, the spatial correlation features of the center wavelength data of each fiber core of the multi-core fiber grating can be extracted, and its dynamic features changing over time can be captured using LSTM layers. On this basis, an attention mechanism is introduced to perform weighted calculations on the hidden states at different time steps, enabling the model to automatically focus on the most representative information decoupled from temperature, vibration, and three-dimensional deformation, thereby enhancing the feature representation ability and learning accuracy.

[0113] Preferably, during model training, the root mean square error (RMSE) is used as the loss function; the Adam optimizer is used to update the network weights through backpropagation algorithm, and the learning rate is set to 0.001-0.01; an early stopping mechanism is set, and training is stopped when the validation set loss does not decrease for 5-10 consecutive rounds to prevent overfitting.

[0114] Specifically, to quantitatively evaluate the error between the model output and the actual physical quantity, the root mean square error (RMSE) is defined as the loss function, and its mathematical expression is as follows:

[0115]

[0116] in, The total number of samples, For the first The true value of each sample The predicted values ​​are used by the model. In each training iteration, the RMSE value is calculated, and the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, which serves as the direction for optimization.

[0117] The adaptive moment estimator (Adam) optimizer is used to update the network weights. The Adam optimizer combines the advantages of momentum gradient descent and adaptive learning rate, and can automatically adjust the learning rate on different parameter dimensions, thereby accelerating model convergence and preventing oscillations.

[0118] During training, the maximum number of iterations is set to [number]. Furthermore, an early stopping mechanism is introduced to prevent overfitting. Specifically, a validation set is set up, and the loss value of the validation set is calculated after each iteration. If continuous... If the loss on the validation set does not show a decreasing trend, the model is considered to have reached its optimal state, training is automatically stopped, and the model parameters that performed best on the validation set are saved.

[0119] Through the above training and optimization steps, the model can achieve stable convergence in complex fiber multi-parameter coupling environments, significantly reducing prediction errors. At the same time, the early stopping mechanism effectively suppresses overfitting, ensuring the model's generalization performance and stability in multi-parameter decoupling tasks involving temperature, vibration, and three-dimensional deformation.

[0120] Preferably, the following evaluation metrics are used to evaluate model performance:

[0121] The coefficient of determination is used to evaluate the goodness of fit of a model to the data.

[0122] Root mean square error (RMSE) measures the degree of deviation between the predicted value and the actual value.

[0123] Average deviation is used to assess the systematic bias in predictions.

[0124] Specifically, to comprehensively evaluate the model's fitting ability and prediction accuracy, the following three evaluation metrics are selected:

[0125] The coefficient of determination, used to evaluate the goodness of fit of a model to the data, is calculated using the following formula:

[0126]

[0127] Root mean square error (RMSE) measures the deviation between predicted and actual values. The formula for calculating it is:

[0128]

[0129] The average deviation, used to assess the systematic bias in forecasts, is calculated using the following formula:

[0130]

[0131] in, The total number of samples, For the first The true value of each sample These are the model's predicted values. This is the mean of the true values.

[0132] By comprehensively evaluating the model using three indicators—coefficient of determination, root mean square error, and mean deviation—the prediction accuracy and stability of the model in multi-parameter coupled problems can be fully reflected.

[0133] Step 5: Input the center wavelength data of the multi-core fiber grating to be tested into the trained CNN-LSTM model, and output the decoupled temperature, vibration and three-dimensional deformation data.

[0134] Specifically, in the model application stage, the preprocessed and normalized center wavelength data of the multi-core fiber grating under test is input into the trained CNN-LSTM model. The model extracts spatial and temporal features through convolutional and recurrent layers, and uses an attention mechanism to focus on information at key moments. Finally, the corresponding predicted values ​​of temperature, vibration, and three-dimensional deformation are generated in the output layer. After inverse normalization, the actual physical quantities can be obtained, realizing fast and accurate decoupling and real-time measurement of multiple parameters.

[0135] Compared with existing technologies, this invention realizes an end-to-end decoupling modeling process for multi-parameter signals, significantly reducing the error accumulation caused by multi-stage fitting in traditional methods, and improving decoupling accuracy and algorithm stability. By introducing attention mechanisms and data augmentation strategies, this method can still maintain excellent noise resistance and generalization ability under complex working conditions, and has strong engineering application and promotion value.

[0136] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-parameter decoupling method for multi-core fiber gratings based on improved CNN-LSTM, characterized in that, include: Step 1: Build a multi-core fiber grating temperature, vibration, and three-dimensional deformation detection system, and acquire the center wavelength data when subjected to temperature, vibration, and three-dimensional deformation respectively, and generate a supervised learning dataset; Step 2: Preprocess the collected center wavelength data of the multi-core fiber grating, and use the MinMax normalization method to normalize the center wavelength data of each fiber core to the [0,1] interval; Step 3: Divide the normalized center wavelength data into training and test sets in a 7:3 ratio, and construct the training dataset by combining stratified sampling and data augmentation. Step 4: Build and train the improved CNN-LSTM network model, and evaluate the test dataset using the trained CNN-LSTM model; Step 5: Input the center wavelength data of the multi-core fiber grating to be tested into the trained CNN-LSTM model, and output the decoupled temperature, vibration and three-dimensional deformation data.

2. The multi-core fiber optic grating multi-parameter decoupling method based on improved CNN-LSTM according to claim 1, characterized in that, The detection system consists of a multi-core fiber grating, a fan-in / fan-out module, a fiber grating demodulator, a heating device, a thermistor, a temperature measuring instrument, a vibration platform, an accelerometer, an acceleration measuring instrument, a standard bending mold, and a cantilever beam.

3. The multi-core fiber optic grating multi-parameter decoupling method based on improved CNN-LSTM according to claim 2, characterized in that, The steps to obtain a supervised learning dataset include: A temperature load is applied to the cantilever beam by a heating device, and a temperature sensor collects temperature change data in real time as the true temperature value. A vibration load is applied to the cantilever beam by a vibration excitation device, and an acceleration sensor collects the vibration acceleration data in real time as the true value of the vibration. A three-dimensional deformation load with a preset curvature is applied to a multi-core fiber optic grating using a standard bending die, and the true value of the three-dimensional deformation is obtained based on the geometric parameters of the die. The center wavelength data of each fiber core of the multi-core fiber grating is collected synchronously under various loads. The center wavelength data is then paired with the corresponding true values ​​of temperature, vibration, and three-dimensional deformation to construct a supervised learning dataset.

4. The multi-core fiber optic grating multi-parameter decoupling method based on improved CNN-LSTM according to claim 1, characterized in that, The preprocessing steps for center wavelength data include: Outlier detection and removal are performed on the collected center wavelength data; The center wavelength data is filtered to remove high-frequency noise and interference signals.

5. The multi-core fiber optic grating multi-parameter decoupling method based on improved CNN-LSTM according to claim 1, characterized in that, The steps to construct the training dataset include: A stratified sampling method was adopted to stratify the data according to different working condition ranges of temperature, vibration and deformation, ensuring that the data of each stratum were distributed proportionally to the training set and the test set. Data augmentation is performed on the training set. Augmentation methods include time shifting, amplitude scaling, adding Gaussian noise, or time warping to increase the number of training samples.

6. The multi-core fiber optic grating multi-parameter decoupling method based on improved CNN-LSTM according to claim 1, characterized in that, The CNN-LSTM network model includes an input layer, hidden layers, attention layers, fully connected layers, and an output layer, wherein: The input layer receives the normalized center wavelength data; The hidden layer includes a CNN convolutional layer and an LSTM network layer. The CNN convolutional layer is used to extract spatial features, and the LSTM network layer extracts temporal features through a gating mechanism. The attention layer calculates attention weights for the hidden state sequence output by the LSTM network layer; The fully connected layer integrates the weighted features; The output layer outputs the decoupled temperature value, vibration value, and three-dimensional deformation value.

7. A multi-core fiber optic grating multi-parameter decoupling method based on an improved CNN-LSTM according to claim 6, characterized in that, The attention weight calculation process for the attention layer includes: Construct a matching scoring function for the hidden state at each time step of the LSTM network layer output and the hidden state of the output layer; The scoring results are normalized using the softmax function to obtain the attention weight vector.

8. The multi-core fiber optic grating multi-parameter decoupling method based on improved CNN-LSTM according to claim 1, characterized in that, During model training, the root mean square error (RMSE) is used as the loss function; the Adam optimizer is used to update the network weights through backpropagation algorithm, and the learning rate is set to 0.001-0.01; an early stopping mechanism is set to stop training when the validation set loss does not decrease for 5-10 consecutive rounds to prevent overfitting.

9. A multi-core fiber optic grating multi-parameter decoupling method based on an improved CNN-LSTM according to claim 1, characterized in that, The following evaluation metrics were used to assess model performance: The coefficient of determination is used to evaluate the goodness of fit of a model to the data. Root mean square error (RMSE) measures the degree of deviation between the predicted value and the actual value. Average deviation is used to assess the systematic bias in predictions.

Images