Search

CN-122023834-A - Multi-scale fiber bragg grating multimodal demodulation method integrating attention mechanism

CN122023834ACN 122023834 ACN122023834 ACN 122023834ACN-122023834-A

Abstract

The invention discloses a multiscale fiber grating multimodal demodulation method integrating an attention mechanism, which comprises the following steps of firstly carrying out pretreatment such as normalization, noise filtering and baseline correction on an original reflection spectrum, secondly carrying out multiscale feature extraction through continuous wavelet transformation, a one-dimensional convolutional neural network and global context analysis, then screening and enhancing features by applying a dual attention mechanism formed by channel attention and space attention, fusing the processed multiscale features, and finally carrying out self-adaptive peak detection and post-processing based on the fused features to realize accurate positioning of central wavelength. The method effectively solves the demodulation difficulty of the traditional method when the optical fiber grating spectrums are overlapped, can still realize near zero-error high-precision demodulation especially in the extreme scene of complete overlapping, has strong anti-interference performance and high robustness, and is suitable for the optical fiber grating sensing network with high density and complex working conditions.

Inventors

  • WANG YINGZHI
  • WANG FEI
  • LI DONGYUE
  • ZHANG CHAO
  • SUN JIAYU

Assignees

  • 长春理工大学
  • 吉林省博辉科技有限公司

Dates

Publication Date
20260512
Application Date
20260202

Claims (8)

  1. 1. A multi-scale fiber grating multi-peak demodulation method integrating an attention mechanism is characterized by comprising the following steps: step 1, carrying out data preprocessing on the multimodal reflection spectrum data of the fiber grating; step 2, extracting multi-scale characteristics of the preprocessed spectrum data; Step 3, processing the extracted multi-scale features by using a dual-attention mechanism; step 4, fusing the multiscale characteristics processed by the attention mechanism; Step 5, performing self-adaptive peak detection based on the fusion characteristics; step 6, post-processing the detected peak value; The multi-scale feature extraction comprises multi-scale time-frequency feature extraction through continuous wavelet transformation, local feature extraction through a one-dimensional convolutional neural network and global statistical feature extraction through global context analysis, wherein the dual-attention mechanism comprises a channel attention mechanism and a space attention mechanism.
  2. 2. The method according to claim 1, wherein the data preprocessing in step 1 specifically includes: Normalizing the original reflection spectrum signal according to the maximum value to compress the reflectivity value range to the [0,1] interval, unifying all spectrum data to the same wavelength range, and adopting the formula: ; Where x is the raw data, x' is the normalized result value, And Respectively the minimum value and the maximum value of the original data; Noise filtering processing, namely noise suppression is carried out by adopting moving average filtering, and a sliding filtering algorithm is as follows: ; Wherein: { x1, x2,., xn } is the point where the sliding mean filter process was performed and its neighboring n-1 data points; baseline correction-estimating background baseline using polynomial fitting, and obtaining corrected spectra by subtracting the fitted baseline.
  3. 3. The method according to claim 1, wherein the multi-scale feature extraction in step 2 specifically comprises: Continuous wavelet transformation feature extraction, namely continuous wavelet transformation is carried out on a plurality of different scales, the mode maximum value point under each scale is extracted, a multi-scale time-frequency feature diagram is constructed, and the continuous wavelet transformation feature extraction formula is as follows: ; ; Wherein, the Is a mother wavelet function, a is a scale parameter, b is a shift parameter corresponding to the wavelength of the spectrum, Is a normalization factor, representing the complex conjugate of the mother wavelet function, Is the spectrum corresponding intensity; The one-dimensional convolutional neural network feature extraction, namely, adopting a plurality of convolutional kernels with different sizes to carry out parallel processing, and constructing multistage features through different-level convolutional operations; ; Where, represents a one-dimensional convolution operation, N is the number of input channels, And The weight and bias of the kth convolution kernel, respectively, f is the activation function; Global context feature extraction, namely carrying out global average pooling on the feature map to obtain the integral statistical feature of the spectrum, and calculating the mean value, variance, skewness and kurtosis statistics of the spectrum.
  4. 4. The method according to claim 1, wherein the dual attention mechanism in step 3 specifically comprises: The channel attention mechanism is used for carrying out global average pooling on each characteristic channel to obtain a channel statistics descriptor, and learning the importance weight of each channel through a fully connected network to realize the characteristic selection on the channel dimension; The space attention mechanism is used for carrying out maximum pooling and average pooling on the feature map along the channel dimension to generate a space descriptor, generating a space attention weight map through a convolution layer and an activation function, and enhancing the feature response of the key region; Double attention fusion, namely multiplying the channel attention and the space attention weight element by element, and recalibrating the original characteristics by applying the fused attention weight.
  5. 5. The method according to claim 1, wherein the multi-scale feature fusion in the step 4 specifically includes unifying all feature maps to the same size through upsampling or downsampling, splicing in a channel dimension, designing a feature interaction module to promote information flow between features of different scales, controlling information flow in a feature fusion process by using a gating mechanism, performing layer normalization on the fused features, and compressing the feature maps to a fixed size by using adaptive pooling.
  6. 6. The method according to claim 1, wherein the adaptive peak detection in step 5 specifically comprises: Coarse positioning of candidate peaks, namely detecting local maximum points on the fusion characteristic response graph, setting a self-adaptive threshold value to filter weak response candidate peaks, and eliminating redundant detection through neighborhood non-maximum suppression; Modeling reflection peaks by adopting Gaussian functions with different left and right standard deviations, initializing fitting parameters based on candidate peak positions and characteristic responses, and performing nonlinear least square fitting by using an optimization algorithm; , Wherein A is the height of the peak, μ is the center position of the peak, σl and σr are the width parameters on the left and right sides of the center, respectively; The confidence coefficient evaluation system is used for calculating the root mean square error of the fitting residual error to evaluate the fitting quality, checking the consistency of the detection peak on the multi-scale characteristics, integrating a plurality of factors to calculate confidence coefficient scores, and setting a dynamic threshold value according to the confidence coefficient scores to filter low confidence coefficient detection.
  7. 7. The method according to claim 1, wherein the post-processing in the step 6 specifically includes realizing sub-pixel level peak positioning by a parabolic interpolation method, calculating a center wavelength, and outputting a detection result.
  8. 8. The method according to any one of claims 1 to 7, further comprising: And seventhly, realizing online self-adaptive optimization of the algorithm through a feedback optimization mechanism, wherein the online self-adaptive optimization comprises confidence feedback adjustment and parameter self-adaptive update.

Description

Multi-scale fiber bragg grating multimodal demodulation method integrating attention mechanism Technical Field The invention relates to the technical field of fiber bragg grating sensors, in particular to a multi-scale fiber bragg grating multimodal demodulation method integrating an attention mechanism. Background Fiber Bragg Grating (FBG) sensors are widely applied to the fields of aerospace, civil engineering, energy power and the like due to the advantages of electromagnetic interference resistance, corrosion resistance, easiness in networking and the like. The FBG sensor senses the change of physical quantities such as temperature, strain and the like through the drift of the reflection center wavelength, so the wavelength demodulation precision directly determines the performance index of the sensing system. The traditional demodulation algorithm mainly comprises a centroid method, a polynomial fitting method, a Gaussian fitting method and the like, wherein the methods can achieve higher precision under ideal spectrum conditions, but face serious challenges in practical engineering application, namely firstly, reflection spectrum signal-to-noise ratio is reduced due to factors such as unstable light source, transmission link loss and environmental noise, secondly, a wavelength division multiplexing technology adopted for realizing multipoint measurement is easy to cause reflection spectrum overlapping to form a complex multimodal structure, and further, spectrum distortion is caused by factors such as sensor encapsulation, structure nonuniform stress and the like, and ideal Gaussian line type is destroyed. Although the existing improved algorithm can improve the precision to a certain extent, the problems of poor self-adaptation capability, weak anti-interference capability, limited demodulation precision and the like still exist when the serious overlapped peak, the distorted peak and the spectrum with low signal to noise ratio are processed. In recent years, although research attempts are made to introduce deep learning technology, most network models have the defects of single feature extraction and insufficient cooperative utilization of spectrum local and global features, and the requirements of high precision and self-adaptive demodulation of multimodal reflection spectrum under complex working conditions are difficult to meet. Therefore, developing a demodulation algorithm which can effectively integrate multi-scale features, is adaptive to a focusing key spectrum region and has strong noise immunity has become a key for promoting further development of fiber grating sensing technology. Disclosure of Invention The technical scheme for solving the technical problems is that the invention provides a multi-scale fiber bragg grating multi-peak demodulation method integrating an attention mechanism, which comprises the following steps: step 1, carrying out data preprocessing on the multimodal reflection spectrum data of the fiber grating; step 2, extracting multi-scale characteristics of the preprocessed spectrum data; Step 3, processing the extracted multi-scale features by using a dual-attention mechanism; step 4, fusing the multiscale characteristics processed by the attention mechanism; Step 5, performing self-adaptive peak detection based on the fusion characteristics; step 6, post-processing the detected peak value; The multi-scale feature extraction comprises multi-scale time-frequency feature extraction through continuous wavelet transformation, local feature extraction through a one-dimensional convolutional neural network and global statistical feature extraction through global context analysis, wherein the dual-attention mechanism comprises a channel attention mechanism and a space attention mechanism. Further, the data preprocessing in the step 1 specifically includes: Normalizing the original reflection spectrum signal according to the maximum value to compress the reflectivity value range to the [0,1] interval, unifying all spectrum data to the same wavelength range, and adopting the formula: ; Where x is the raw data, x' is the normalized result value, AndRespectively the minimum value and the maximum value of the original data; Noise filtering processing, namely noise suppression is carried out by adopting moving average filtering, and a sliding filtering algorithm is as follows: ; Wherein: { x1, x2,., xn } is the point where the sliding mean filter process was performed and its neighboring n-1 data points; baseline correction-estimating background baseline using polynomial fitting, and obtaining corrected spectra by subtracting the fitted baseline. Further, the multi-scale feature extraction in step 2 specifically includes: Continuous wavelet transformation feature extraction, namely continuous wavelet transformation is carried out on a plurality of different scales, the mode maximum value point under each scale is extracted, a multi-scale time-frequency feature diagram is construc