Fiber bragg grating fast demodulation method based on hybrid learning framework

Abstract

The application belongs to the technical field of fiber Bragg grating demodulation, and specifically discloses a fiber Bragg grating rapid demodulation method based on a hybrid learning framework, which comprises the following steps: acquiring a reflection spectrum sequence of a fiber Bragg grating; inputting the sequence into a TimeMixer model of a pure multi-layer perception machine architecture, performing average pooling downsampling to obtain a multi-scale sequence, projecting the sequence of each scale into a deep feature, mixing the deep features of different scales through a stacked past decomposition mixing module to obtain multi-scale past information, and then inputting the multi-scale past information into a multi-future prediction mixing module, using an independent predictor to predict each scale, and intercepting the deep feature output by the prediction layer as a spectrum sequence feature; splicing the features of each scale to obtain a combined feature; and inputting the combined feature into a LightGBM model to regress to obtain a center wavelength prediction value. The application can realize a millisecond-level fast response while maintaining a picometer-level precision.

Description

Technical Field

[0001] This application belongs to the field of fiber Bragg grating demodulation technology, and more specifically, relates to a fast demodulation method for fiber Bragg gratings based on a hybrid learning framework. Background Technology

[0002] Fiber Bragg grating (FBG) demodulation technology is widely used in many monitoring fields due to its advantages such as high precision, resistance to electromagnetic interference, intrinsic safety, and ease of multiplexing and networking. However, long-term operation of FBGs under environmental factors such as radiation and high temperature can easily cause spectral distortion and noise increase, thereby affecting spectral stability and leading to distorted reflection spectra. In addition, light source fluctuations and interference noise in the demodulation system can further degrade signal quality. Therefore, achieving high-precision demodulation of distorted spectra is crucial to ensuring the reliability and stability of grating sensor networks.

[0003] To address the spectral distortion problem of FBG reflectance, existing optimization techniques can be divided into hardware and software aspects. Hardware optimization mainly involves improving manufacturing processes, grating structure design, and optimizing packaging and installation, while software optimization focuses on enhancing signal processing and demodulation algorithms. However, when FBGs are deployed in complex environments (such as embedded in concrete structures, implanted in the human body, or buried deep in oil wells) or when the packaging process has inherent defects, physically modifying the sensor is often not feasible. In contrast, signal processing algorithms can directly post-process the acquired distorted spectrum without replacing the sensor or interrupting the system, offering significant advantages such as low cost and rapid implementation.

[0004] By uncovering the statistical patterns behind data, machine learning can obtain reliable and repeatable judgments. Traditional machine learning techniques have been applied to FBG demodulation. With the rapid development of artificial intelligence, especially deep learning, it has attracted widespread attention in the field of fiber optic sensing demodulation due to its powerful nonlinear feature extraction capabilities. Existing research has proposed deep learning methods based on Long Short-Term Memory (LSTM) networks, LSM convolutional neural networks, one-dimensional convolutional neural networks, convolutional neural network-LSTM network architectures, and temporal convolutional networks, achieving effective identification and demodulation of distorted spectra. Although the above deep learning models have shown significant accuracy advantages in FBG distorted spectrum demodulation, this advantage often comes at the cost of complex network structures and huge computational overhead. For example, the convolutional computation and multi-layer stacking of convolutional neural networks, and the sequence dependencies and complex gating of LSM networks, all significantly increase computational latency. Even with the introduction of parallel computing in temporal convolutional networks, multiple layers of dilated convolutions are often required to ensure full-spectrum feature coverage, which not only increases the number of parameters but also limits the achievement of extreme inference speed. In time-sensitive applications such as earthquake monitoring, high-speed rotating machinery condition monitoring, and real-time load analysis of aerospace vehicles, traditional deep learning frameworks struggle to meet millisecond-level response requirements, and the bottleneck of demodulation speed severely restricts the system's practicality.

[0005] Therefore, how to achieve a fast response at the millisecond or even microsecond level while demodulating the distortion spectrum with high precision is a problem that urgently needs to be solved. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this application is to provide a fast demodulation method for fiber Bragg gratings based on a hybrid learning framework, which can achieve millisecond-level or even microsecond-level fast response while demodulating distorted spectra with high precision.

[0007] To achieve the above objectives, in a first aspect, this application provides a fast demodulation method for fiber Bragg gratings based on a hybrid learning framework, comprising the following steps: S10, Obtain the reflection spectrum sequence of the fiber Bragg grating; S20, the reflectance spectral sequence is input into the TimeMixer model of a pure multilayer perceptron architecture, and the following operations are performed: the reflectance spectral sequence is downsampled by average pooling to obtain a multi-scale sequence; the sequence of each scale is projected into depth features through an embedding layer; the depth features of different scales are mixed through a stacked past decomposition mixing module to obtain multi-scale past information; the multi-scale past information is input into a multi-future prediction mixing module, wherein an independent predictor is used for prediction at each scale, each predictor contains a prediction layer and a projection layer, and the depth features output by the prediction layer of each predictor are extracted as the spectral sequence features of that scale; S30, the spectral sequence features at each scale are spliced ​​together to obtain combined features; S40, the combined features are input into the LightGBM model, and the center wavelength prediction value is obtained through gradient boosting decision tree regression.

[0008] As a further preferred embodiment, in step S20, the average pooling downsampling and projection specifically include: Average pooling is used for the input signal Perform downsampling to M A set of multi-scale sequences is obtained by considering multiple scales. , No. m Sequences of several scales ,in P For sequence length, C For the number of variables, It is an input sequence containing the finest spectral details. Reflects changes in the macroscopic profile of the spectrum; Multi-scale sequences are projected onto deep features through an embedding layer. In, it is represented as .

[0009] As a further preferred embodiment, in step S20, the mixing of depth features at different scales through stacked past decomposition mixing modules specifically includes: For the The layered decomposition and mixing module, the input is The process of decomposing hybrid modules in the past can be formalized into a formula: ,in It is the total number of floors. This represents the preorder feature representation after mixing; The past decomposition and mixing module decomposes the sequence into... Decomposed into seasonal components and trend section Trend item Corresponding spectral baseline and overall profile, seasonal term It captures high-frequency local details of the spectrum, including side lobes and noise, and uses a bottom-up mixing method for seasonal terms and a top-down mixing method for trend terms.

[0010] As a further preferred embodiment, the bottom-up mixing approach for the seasonal terms specifically involves integrating fine-grained sequence information from lower levels upwards, i.e., the first... m Bottom-up mixing at each scale according to the formula Execution, in which S m For the first m Seasonal items at a certain scale, Two linear layers along the time dimension are used, with the GELU activation function in between; The aforementioned top-down hybrid approach to trend items specifically involves: using macroscopic knowledge from coarser-grained data to guide trend modeling of finer-grained data. m Top-down mixing at each scale according to the formula Execution, in which t m For the first m Trend terms at each scale, It employs two linear layers along the time dimension, with the GELU activation function in between.

[0011] As a further preferred embodiment, in step S20, the process by which the multiple future prediction hybrid module uses an independent predictor for each scale to make predictions specifically is as follows: Let the multi-scale past information obtained after passing through L past decomposition and mixing modules be . For the first m Each scale uses an independent predictor. right To make predictions, the prediction layer of the predictor follows the formula. A deep representation of the future sequence regressed from past information at this scale. , The predictor's projection layer is calculated according to the formula. The output of the prediction layer A depth representation projected onto Given several target variable dimensions, generate the final prediction at that scale. The process of extracting the depth features from the prediction layer output of each predictor is called truncation. .

[0012] As a further preferred embodiment, in step S40, the LightGBM model employs a histogram algorithm to discretize continuous feature values ​​into a finite number of k buckets, and constructs a histogram for each feature to statistically analyze the gradient information of samples within the bucket. When searching for the optimal split point, these histogram buckets are traversed, reducing the time complexity from the number of features × the number of samples to the number of features × the number of histogram buckets. A one-sided gradient sampling algorithm is used, prioritizing the retention of samples with large gradients and randomly sampling samples with small gradients, assigning weight compensation to the sampled small gradient samples to maintain the unbiasedness of the data distribution. A mutually exclusive feature bundling algorithm is used to bundle multiple mutually exclusive features into a new feature, compressing the dimension of the feature space from the original number of features to the number of bundled features, where the number of bundled features is much smaller than the original number of features. A depth-constrained leaf-wise growth strategy is adopted, selecting the leaf node with the largest splitting gain from all current leaf nodes in each iteration for splitting, and preventing overfitting by limiting the maximum depth.

[0013] As a further preferred embodiment, in step S10, the reflectance spectrum sequence is generated in the following manner: The simulation results were obtained using an improved model based on the asymmetric generalized Gaussian function. The mathematical expression of this model is given by the formula:

[0014] In the formula, λ Represents the wavelength variable of the reflectance spectrum; It is at the Prague wavelength Maximum reflectivity at the location; morphological factor Determines the sharpness of the peak value. The larger the value, the sharper the peak; asymmetry factor Adjusting the spectral skewness characteristics, when The spectrum exhibits a left-skewed distribution, when If the spectrum exhibits a right-skewed distribution, then the spectrum is tilted to the left or right.

[0015] As a further preferred embodiment, the reflectance spectral sequence is a double-peaked overlapping spectrum, which is obtained by linearly superimposing two independently generated fiber Bragg grating reflectance spectra and adding additive Gaussian white noise, as expressed by the formula: Each single-peak spectrum An improved model based on the asymmetric generalized Gaussian function is used for generation. Additive white Gaussian noise; each bimodal sample contains two Bragg wavelength labels. And satisfy .

[0016] As a further preferred embodiment, the double-peak overlapping spectrum is divided into three overlap levels according to the range of the center wavelength spacing Δλ: slight overlap is when Δλ belongs to the range of 0.30 nm to 0.50 nm, the two peaks are close but there are distinguishable peak valleys; moderate overlap is when Δλ belongs to the range of 0.15 nm to 0.30 nm, the two peaks are merged and traditional peak finding methods fail; severe overlap is when Δλ belongs to the range of 0.02 nm to 0.15 nm, the two peaks are merged into a single peak, and can be distinguished only by the asymmetry of the spectrum; each overlap level generates a sample independently, each spectral sample contains 2001 data points, the wavelength range covers 1549 nm to 1551 nm, and the step size is 0.001 nm.

[0017] Secondly, this application provides a fiber Bragg grating fast demodulation system for implementing the method described in any one of the above, comprising: The spectrum acquisition module is used to acquire the reflection spectrum sequence of the fiber Bragg grating; The feature extraction module internally deploys a TimeMixer model to perform average pooling downsampling on the reflected light sequence to obtain a multi-scale sequence. The sequence at each scale is projected into depth features through an embedding layer. The depth features at different scales are mixed through a stacked past decomposition mixing module to obtain multi-scale past information. The multi-scale past information is input into a multi-future prediction mixing module, which uses an independent predictor for each scale. Each predictor contains a prediction layer and a projection layer. The depth features output by the prediction layer of each predictor are extracted as spectral sequence features at each scale, and the spectral sequence features at each scale are concatenated into a combined feature. The wavelength regression module, which internally deploys the LightGBM model, is used to obtain the predicted value of the center wavelength by regressing the combined features through gradient boosting decision tree.

[0018] This application offers the following advantages: By employing a TimeMixer model with a pure multilayer perceptron architecture, average pooling downsampling is used to obtain multi-scale sequences from the reflectance spectral sequences. Its multi-scale mixing mechanism efficiently decouples and extracts global contour information and local fine features from the signal. Furthermore, the core operation of the pure multilayer perceptron architecture is highly parallel matrix multiplication, which offers higher computational efficiency at the hardware level compared to convolution operations in traditional convolutional neural networks, laying a hardware foundation for improved inference speed. The deep features output from the prediction layers at each scale are concatenated and input into the LightGBM model. Its gradient boosting tree robustly maps high-dimensional spectral features, and histogram acceleration algorithms and leaf-wise growth strategies enable ultra-high-speed wavelength shift prediction. Therefore, this application achieves millisecond- or even microsecond-level single-spectrum processing speeds while maintaining picometer-level demodulation accuracy, significantly improving inference speed compared to existing deep learning schemes, and exhibiting excellent robustness to noisy environments. Attached Figure Description

[0019] Figure 1 This is the model architecture of TimeMixer provided in this application; Figure 2 This application provides a time-linear layer for mixing seasonal terms; Figure 3 This application provides a time-linear layer for mixing trend terms; Figure 4 The decision tree growth strategy provided in this application is (a) a Leaf-wise growth strategy and (b) a Level-wise growth strategy. Figure 5 This is the overall flow of the TimeMixer-LightGBM model provided in this application; Figure 6This application provides a TimeMixer feature extraction and LightGBM wavelength prediction framework. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] To overcome the dual constraints of real-time performance and accuracy, this application proposes a TimeMixer-LightGBM hybrid demodulation framework. The TimeMixer model employs a pure multilayer perceptron architecture, with its core operation being highly parallel matrix multiplication. Compared to convolution operations in traditional convolutional neural networks, it boasts higher computational efficiency at the hardware level, laying a hardware foundation for improved inference speed. Through its unique multi-scale mixing mechanism, TimeMixer can efficiently decouple and extract global contour information and local fine features from signals. This application innovatively introduces advanced architectures from the field of time series analysis into spectral analysis, treating FBG reflectance spectra as one-dimensional sequence data and utilizing TimeMixer to extract multi-scale spectral sequence features. Subsequently, LightGBM is used for rapid prediction of the center wavelength. The LightGBM regressor uses gradient boosting trees to robustly map high-dimensional spectral features, and its histogram acceleration algorithm and leaf-wise growth strategy ensure ultra-high-speed wavelength shift prediction. Experiments show that while maintaining picometer-level demodulation accuracy (RMSE = 2.128 pm), the framework reduces the single-spectrum processing time to 0.08 milliseconds, which is more than 4.5 times faster than existing deep learning solutions, providing a new paradigm for real-time dynamic monitoring of FBG sensor networks.

[0022] (1) TimeMixer-LightGBM hybrid learning demodulation framework TimeMixer is a sequence prediction model based purely on a multilayer perceptron architecture. This model significantly improves the accuracy and efficiency of sequence analysis through innovative multi-scale decomposition and mixing mechanisms. This application incorporates FBG reflectance spectra as one-dimensional sequence data into the model. Figure 1 The model architecture of TimeMixer is demonstrated. The model first decomposes the input spectral sequence into multiple scales; then, it projects these multiple scale sequences into deep features through embedding techniques; next, multiple scale mixed predictions are achieved by two main modules: Past-Decomposable-Mixing (PDM) and Future-Multipredictor-Mixing (FMM).

[0023] To obtain a multi-scale spectral sequence, TimeMixer uses average pooling to downsample the input signal to M scales, resulting in a set of multi-scale sequences, as shown in formula (1): (1) in , Indicates the number of variables. It is an input sequence containing the finest spectral details. This reflects changes in the macroscopic contours of the spectrum. Then, these multi-scale sequences are projected onto deep features through an embedding layer. In, it is represented as Through the above design, we obtained a multi-scale representation of the input sequence.

[0024] Next, stacked PDM blocks are used to mix prior spectral information at different scales (corresponding to "past" information in the time series model). For the first... Layer, input is The PDM process can be formalized as formula (2): (2) in, It is the total number of floors. This represents the preorder feature representation after mixing.

[0025] The PDM module decomposes multi-scale time series data. Break it down into seasonal components. And the trend section (Trend) Under spectral analysis, the trend term Corresponding to the baseline and overall profile of the spectrum, while the seasonal term This captures high-frequency local details such as sidelobes and noise in the spectrum. Considering the different natures of seasonality and trend, a hybrid operation is applied to both to integrate multi-scale information.

[0026] In seasonal item mixing, a bottom-up approach is used, such as... Figure 2 As shown, fine-grained sequence information from lower levels is integrated upwards to supplement more detailed information for coarser-scale seasonal modeling. The bottom-up mixing at each scale can be formally represented by formula (3): (3) in It employs two linear layers along the time dimension, with the GELU activation function in between.

[0027] In trend item blending, a top-down blending method is used, such as... Figure 3As shown, macroscopic knowledge from coarser-grained data is used to guide trend modeling of finer-grained data. The top-down mixing at the m-th scale can be formally expressed as formula (4): (4) in It employs two linear layers along the time dimension, with the GELU activation function in between.

[0028] After L PDM blocks, we obtain multi-scale past information. Next, the multi-scale sequences are predicted and aggregated using FMM blocks. The FMM process is as follows: First, there is multi-scale prediction, for each scale... Using a separate predictor Regarding past information The prediction is formally represented by formula (5): (5) in, This represents the subsequent spectral predictions generated from the m-th scale sequence.

[0029] The prediction results are then integrated by weighted summation of predictions from all scales to obtain the final prediction result. Through this design, TimeMixer can fully utilize the complementary prediction capabilities of sequences at different scales, thereby improving overall prediction performance.

[0030] In this application, LightGBM is a high-efficiency machine learning algorithm based on gradient boosting decision trees. Compared to traditional GBDT implementations, LightGBM has undergone deep optimization in terms of training speed and memory consumption, making it particularly suitable for processing high-dimensional data and real-time application scenarios. This framework chooses LightGBM to leverage its extremely fast inference speed to quickly and robustly map the deep features extracted by TimeMixer to the center wavelength value. Its core optimization techniques include: To accelerate the search for split points in decision trees, LightGBM employs a histogram algorithm instead of the traditional pre-sorting algorithm. This algorithm first discretizes continuous feature values ​​into a finite number (k) of buckets and constructs a histogram for each feature to statistically analyze the gradient information of samples within each bucket. When searching for the optimal split point, the algorithm only needs to traverse these histogram buckets, rather than all sample data, thus reducing the time complexity from... Significantly reduced to At the same time, it significantly reduces memory access overhead, which is the key to its high computing efficiency.

[0031] One-sided Gradient Sampling (GOSS) algorithm improves efficiency by optimizing the training sample set. Its core idea is that samples with smaller gradients during model training are usually well-learned and contribute less to gain calculation. Therefore, GOSS prioritizes retaining samples with larger gradients and only randomly samples samples with smaller gradients. Simultaneously, it assigns weights to these small-gradient samples to maintain the unbiasedness of the data distribution. This strategy effectively reduces the computational burden while ensuring model accuracy.

[0032] To address the feature sparsity problem prevalent in high-dimensional data, LightGBM proposes the Mutually Exclusive Feature Bundling (EFB) algorithm. This algorithm bundles multiple mutually exclusive features (those that are rarely simultaneously non-zero) into a new single feature, thereby reducing the dimensionality of the feature space from... Compress to (in This lossless feature reduction technique further reduces the computational load and memory consumption during model training.

[0033] Unlike most GBDT tools that use a level-wise growth strategy, LightGBM employs a depth-constrained leaf-wise growth strategy. For example... Figure 4 As shown in (a), this strategy selects the leaf node with the largest splitting gain from all current leaf nodes for splitting in each iteration, instead of... Figure 4 The level-wise strategy shown in (b) splits all leaves in the same layer. This approach can reduce loss more purposefully, thus achieving higher accuracy with fewer iterations. At the same time, by limiting the maximum depth, overfitting can be effectively prevented.

[0034] The TimeMixer-LightGBM hybrid demodulation framework proposed in this application aims to address the core challenge of balancing high accuracy and low latency in FBG sensing systems. The core idea of ​​this framework is division of labor and collaboration: first, the TimeMixer model is used to extract powerful multi-scale deep features from the FBG reflectance spectrum (self-supervised learning), and then the LightGBM model is used to perform fast and robust final regression (supervised learning). Figure 5 The overall process based on the TimeMixer-LightGBM model is demonstrated, which includes three core steps: dataset establishment and preprocessing, deep feature extraction based on TimeMixer, and center wavelength regression based on LightGBM.

[0035] In the dataset establishment and preprocessing stage, to obtain sufficient FBG reflectance spectral data that closely matches actual working conditions, this application introduces an improved model based on an asymmetric generalized Gaussian function for spectral simulation. This model can flexibly characterize normal and distorted spectra, and its mathematical expression is shown in formula (6): (6) In the formula, This represents the wavelength variation in the FBG spectrum. It is at the Prague wavelength Maximum reflectivity at location, morphological factor Determines the sharpness of the peak value. The larger the value, the sharper the peak. Asymmetry factor. Adjusting spectral skewness characteristics. When The spectrum exhibits a left-skewed distribution, when If the spectrum exhibits a right-skewed distribution, then the spectrum is tilted to the left or right. Different and The values ​​correspond to different degrees of asymmetry. When and ,but This can be simplified to a Gaussian symmetric model. Through coordinated regulation... and It can efficiently generate a series of spectral samples simulating complex distortions, providing a crucial data foundation for demodulation algorithms.

[0036] During data generation, parameters were randomly selected within a set range, and random Gaussian white noise was added to simulate the actual measurement environment. All generated spectral data underwent standardized preprocessing and were proportionally divided into training, validation, and test sets to ensure the effectiveness of model training and the reliability of evaluation.

[0037] In the TimeMixer-based feature extraction and LightGBM-based wavelength prediction stages, this application innovatively uses the TimeMixer model as a deep feature extractor. The multi-scale decomposition and mixing mechanism of TimeMixer is very suitable for processing sequential signals containing complex patterns, such as FBG reflectance spectra.

[0038] In the first stage, the TimeMixer model is trained using the training set. During this stage, TimeMixer's task is spectral sequence prediction / reconstruction, i.e., learning the intrinsic relationships between different wavelengths in the spectrum. The validation set is then used to evaluate model performance and select the optimal model architecture and hyperparameters.

[0039] In the second stage, after verifying its generalization ability on the test set, the parameters of the fully trained TimeMixer model are frozen and used as a fixed feature extractor to obtain the corresponding deep features in batches across the entire dataset.

[0040] The core operation of feature extraction is located in the FMM module of TimeMixer. In this module, each independent predictor at each scale consists of a prediction layer and a projection layer. For the predictor at scale m, its prediction layer is responsible for extracting past information from that scale. Regression to future sequence The process can be formally represented by formula (7): (7) Subsequently, the projection layer will output the prediction layer's output. A depth representation projected onto Given several target variable dimensions, generate the final prediction at that scale. The process can be formally represented by formula (8): (8) Figure 6 This application demonstrates a feature extraction framework using TimeMixer and a wavelength prediction framework using LightGBM. The output of the prediction layer in each predictor is extracted from this application. As a depth feature extracted from the spectral sequence at this scale. All scales The components are concatenated to form the feature input of the LightGBM model. Therefore, the features extracted from the TimeMixer training set data constitute the new training set of LightGBM, while the features extracted from the TimeMixer validation and test sets are correspondingly transformed into the validation and test sets of LightGBM.

[0041] Finally, the reconstructed feature dataset is imported into the LightGBM model for center wavelength prediction. LightGBM, with its histogram algorithm and other optimization techniques, can efficiently process these high-dimensional features generated by TimeMixer, achieving fast and accurate regression of wavelength offsets.

[0042] Experimental comparative analysis revealed that the proposed method was used to demodulate the simulated FBG reflectance spectrum, and the results were compared with existing GAF-CNN-LSTM deep learning methods. As shown in Table 1, in terms of demodulation accuracy, the proposed method achieved a single-spectrum average root mean square error (RMSE) of 2.128 picometers (pm) and a mean absolute error (MAE) of 1.656 picometers (pm), which are on the same picometer scale as the comparative method (RMSE = 2.117 pm, MAE = 1.269 pm). Regarding demodulation real-time performance, the proposed method achieved a single-spectrum average inference time of 0.08 milliseconds (ms), approximately 4.5 times that of the comparative method (0.363 ms). These results demonstrate that the proposed method can achieve microsecond-level fast response while maintaining picometer-level demodulation accuracy, effectively solving the technical challenge of balancing accuracy and real-time performance in FBG demodulation.

[0043] Table 1. Comparison of demodulation performance between TimeMixer-LightGBM and mainstream deep learning models.

[0044] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A fast demodulation method for fiber Bragg gratings based on a hybrid learning framework, characterized in that, Includes the following steps: S10, Obtain the reflection spectrum sequence of the fiber Bragg grating; S20, the reflectance spectral sequence is input into the TimeMixer model of a pure multilayer perceptron architecture, and the following operations are performed: the reflectance spectral sequence is downsampled by average pooling to obtain a multi-scale sequence; the sequence of each scale is projected into depth features through an embedding layer; the depth features of different scales are mixed through a stacked past decomposition mixing module to obtain multi-scale past information; the multi-scale past information is input into a multi-future prediction mixing module, wherein an independent predictor is used for prediction at each scale, each predictor contains a prediction layer and a projection layer, and the depth features output by the prediction layer of each predictor are extracted as the spectral sequence features of that scale; S30, the spectral sequence features at each scale are spliced ​​together to obtain combined features; S40, the combined features are input into the LightGBM model, and the center wavelength prediction value is obtained through gradient boosting decision tree regression.

2. The fast demodulation method for fiber Bragg gratings based on a hybrid learning framework as described in claim 1, characterized in that, In step S20, the average pooling downsampling and projection specifically involve: Average pooling is used for the input signal Perform downsampling to M A set of multi-scale sequences is obtained by considering multiple scales. , No. m Sequences of several scales ,in P For sequence length, C For the number of variables, It is an input sequence containing the finest spectral details. Reflects changes in the macroscopic profile of the spectrum; Multi-scale sequences are projected onto deep features through an embedding layer. In, it is represented as .

3. The fast demodulation method for fiber Bragg gratings based on a hybrid learning framework as described in claim 1, characterized in that, In step S20, the mixing of depth features at different scales through stacked past decomposition mixing modules specifically includes: For the The layered decomposition and mixing module, the input is The process of decomposing hybrid modules in the past can be formalized into a formula: ,in It is the total number of floors. This represents the preorder feature representation after mixing; The past decomposition and mixing module decomposes the sequence into... Decomposed into seasonal components and trend section Trend item Corresponding spectral baseline and overall profile, seasonal term It captures high-frequency local details of the spectrum, including side lobes and noise, and uses a bottom-up mixing method for seasonal terms and a top-down mixing method for trend terms.

4. The fast demodulation method for fiber Bragg gratings based on a hybrid learning framework as described in claim 3, characterized in that, The bottom-up mixing approach for seasonal terms specifically involves integrating fine-grained sequence information from lower levels upwards, i.e., the first... m Bottom-up mixing at each scale according to the formula Execution, in which S m For the first m Seasonal items at a certain scale, Two linear layers along the time dimension are used, with the GELU activation function in between; The aforementioned top-down hybrid approach to trend items specifically involves: using macroscopic knowledge from coarser-grained data to guide trend modeling of finer-grained data. m Top-down mixing at each scale according to the formula Execution, in which t m For the first m Trend terms at each scale, It employs two linear layers along the time dimension, with the GELU activation function in between.

5. The fast demodulation method for fiber Bragg gratings based on a hybrid learning framework as described in claim 1, characterized in that, In step S20, the process by which the multi-future prediction hybrid module uses an independent predictor for each scale to make predictions is as follows: Let the multi-scale past information obtained after passing through L past decomposition and mixing modules be . For the first m Each scale uses an independent predictor. right To make predictions, the prediction layer of the predictor follows the formula. A deep representation of the future sequence regressed from past information at this scale. , The predictor's projection layer is calculated according to the formula. The output of the prediction layer A depth representation projected onto Given several target variable dimensions, generate the final prediction at that scale. The process of extracting the depth features from the prediction layer output of each predictor is called truncation. .

6. The fast demodulation method for fiber Bragg gratings based on a hybrid learning framework as described in claim 1, characterized in that, In step S40, the LightGBM model employs a histogram algorithm to discretize continuous feature values ​​into a finite number of k buckets, and constructs a histogram for each feature to statistically analyze the gradient information of samples within the bucket. When searching for the optimal split point, these histogram buckets are traversed, reducing the time complexity from the number of features × the number of samples to the number of features × the number of histogram buckets. A one-sided gradient sampling algorithm is used, prioritizing the retention of samples with large gradients and randomly sampling samples with small gradients, assigning weight compensation to the sampled small gradient samples to maintain the unbiasedness of the data distribution. A mutually exclusive feature bundling algorithm is used to bundle multiple mutually exclusive features into a new feature, compressing the dimension of the feature space from the original number of features to the number of bundled features, where the number of bundled features is much smaller than the original number of features. A depth-constrained leaf-wise growth strategy is adopted, selecting the leaf node with the largest splitting gain from all current leaf nodes in each iteration for splitting, and limiting the maximum depth to prevent overfitting.

7. The fast demodulation method for fiber Bragg gratings based on a hybrid learning framework as described in claim 1, characterized in that, In step S10, the reflectance spectrum sequence is generated in the following manner: The simulation results were obtained using an improved model based on the asymmetric generalized Gaussian function. The mathematical expression of this model is given by the formula: In the formula, λ Represents the wavelength variable of the reflectance spectrum; It is at the Prague wavelength Maximum reflectivity at the location; morphological factor Determines the sharpness of the peak value. The larger the value, the sharper the peak; asymmetry factor Adjusting the spectral skewness characteristics, when The spectrum exhibits a left-skewed distribution, when If the spectrum exhibits a right-skewed distribution, then the spectrum is tilted to the left or right.

8. The fast demodulation method for fiber Bragg gratings based on a hybrid learning framework as described in claim 1, characterized in that, The reflection spectrum sequence is a double-peak overlapping spectrum, which is obtained by linearly superimposing two independently generated fiber Bragg grating reflection spectra and adding additive Gaussian white noise, as expressed by the formula. Each single-peak spectrum An improved model based on the asymmetric generalized Gaussian function is used for generation. Additive white Gaussian noise; each bimodal sample contains two Bragg wavelength labels. And satisfy .

9. The fast demodulation method for fiber Bragg gratings based on a hybrid learning framework as described in claim 8, characterized in that, The double-peak overlapping spectra are divided into three overlap levels according to the range of the center wavelength spacing Δλ: slight overlap is when Δλ belongs to the range of 0.30 nm to 0.50 nm, the two peaks are close but there are distinguishable peak valleys; moderate overlap is when Δλ belongs to the range of 0.15 nm to 0.30 nm, the two peaks are merged and traditional peak finding methods fail; severe overlap is when Δλ belongs to the range of 0.02 nm to 0.15 nm, the two peaks are merged into a single peak, and can only be distinguished by the asymmetry of the spectrum. Each overlap level generates a sample independently, and each spectral sample contains 2001 data points, covering the wavelength range of 1549 nm to 1551 nm, with a step size of 0.001 nm.

10. A fiber Bragg grating fast demodulation system, characterized in that, To implement the method of any one of claims 1 to 9, comprising: The spectrum acquisition module is used to acquire the reflection spectrum sequence of the fiber Bragg grating; The feature extraction module internally deploys a TimeMixer model to perform average pooling downsampling on the reflected light sequence to obtain a multi-scale sequence. The sequence at each scale is projected into depth features through an embedding layer. The depth features at different scales are mixed through a stacked past decomposition mixing module to obtain multi-scale past information. The multi-scale past information is input into a multi-future prediction mixing module, which uses an independent predictor for each scale. Each predictor contains a prediction layer and a projection layer. The depth features output by the prediction layer of each predictor are extracted as spectral sequence features at each scale, and the spectral sequence features at each scale are concatenated into a combined feature. The wavelength regression module, which internally deploys the LightGBM model, is used to obtain the predicted value of the center wavelength by regressing the combined features through gradient boosting decision tree.

Images