Search

CN-121977683-A - Neural network-assisted triple mixing PGC differential demodulation low-frequency vibration signal detection method

CN121977683ACN 121977683 ACN121977683 ACN 121977683ACN-121977683-A

Abstract

The invention belongs to the technical field of phase demodulation of interference type optical fiber sensors, and discloses a neural network-assisted triple mixing PGC differential demodulation low-frequency vibration signal detection method. The method comprises the steps of constructing a frequency multiplication mixing PGC interference differential sensing system, simulating according to an interference output model to generate a light intensity sequence and a mixing characteristic component, constructing a supervised data set by taking an orthogonal characteristic vector as input and a real-time modulation depth C as a label, training FCNN by adopting a ReLU activation function and Adamax optimizer, evaluating the performance of the model by means of a mean square error and saving qualified weight, inputting the mixing component of the actual sensing system into the model, outputting a smooth C value, substituting the smooth C value into a differential arctangent demodulation algorithm, eliminating the influence of C value fluctuation and restoring a signal to be detected. The invention solves the mathematical singular peak problem of the triple mixing algorithm, maintains the high linearity of the physical model, improves the environment parameter drift resistance and the extremely low frequency signal detection capability of the system, and is suitable for high-stability demodulation of various optical fiber low frequency vibration and acoustic wave sensors.

Inventors

  • ZHANG FAXIANG
  • FENG XINGLONG
  • JIANG SHAODONG
  • LI XINHAO

Assignees

  • 齐鲁工业大学(山东省科学院)

Dates

Publication Date
20260505
Application Date
20260403

Claims (6)

  1. 1. A neural network assisted triple mixing PGC differential demodulation low-frequency vibration signal detection method is characterized by comprising the following steps: S1, acquiring interference signals output by two paths of opposite sensitive interference type optical fiber sensing systems, wherein the interference signals comprise a signal interferometer and a reference interferometer, the signal interferometer and the reference interferometer are arranged in the same closed environment and are opposite to the vibration sensitive direction, and the interference signals comprise phase generation carrier PGC information determined by carrier modulation depth C; S2, extracting features, namely mixing two groups of interference signals with fundamental wave, second harmonic and third harmonic carrier signals respectively, and extracting orthogonal component feature vectors V= [ V 1 ,V 2 ,V 3 ] containing 1-3-order Bessel function items through low-pass filtering; S3, modulating depth fitting based on a neural network, namely directly outputting the real-time carrier modulating depth C at the current moment by taking the orthogonal component V extracted in the step S2 as input through a pre-trained nonlinear mapping relation by using a fully-connected neural network FCNN, wherein a neural network model is formed by utilizing a simulation data set containing C value fluctuation and training by adopting a Adamax optimizer and is used for eliminating mathematical singularity peaks generated by division operation in traditional analytic calculation; And S4, coefficient correction, namely calculating the quadrature component V 1 、V 2 、V 3 extracted in the S2 by utilizing the smooth carrier modulation depth C output in the S3 and combining with a Bessel function recurrence relation to construct a correction coefficient T for phase reduction, wherein the expression is as follows: ; V 1 、V 2 、V 3 is the quadrature components of the first harmonic, the second harmonic and the third harmonic extracted in the step S2; S5, differential restoration and unwrapping, namely carrying out triangular function domain differential processing on demodulation parameters of the signal interferometer and the reference interferometer before arc tangent, counteracting environmental interference by utilizing the characteristic that two paths of common mode noise are approximately equal, obtaining wrapped phases through arc tangent operation, and restoring continuous phase signals to be detected, which are not influenced by carrier modulation depth drift, through a phase unwrapping algorithm.
  2. 2. The neural network-assisted triple-mixing PGC differential demodulation low-frequency vibration signal detection method according to claim 1, wherein the expression of the light intensity signals I output by the signal interferometer and the reference interferometer in S1 is: ; Wherein A is direct current light intensity, B is alternating current term amplitude related to interference fringe contrast, C is carrier modulation depth, omega 0 is carrier angular frequency, phi is phase signal to be detected, and delta phi is common mode noise.
  3. 3. The method for detecting low-frequency vibration signals by using the neural network to assist triple mixing PGC differential demodulation according to claim 2, wherein the orthogonal components extracted by the signal interferometer in S2 are respectively: first harmonic component: ; second harmonic component: ; third harmonic component: ; Wherein J n (C) is an n-th order Bessel function.
  4. 4. The method for detecting a low-frequency vibration signal by using a neural network to assist triple mixing PGC differential demodulation according to claim 3, wherein the constructing process of the neural network in S3 includes: (1) Establishing a multi-layer perceptron MLP comprising an input layer, a hidden layer and an output layer; (2) Nodes of the input layer receive normalized orthogonal component sequences (V1, V2, V3) at the current sampling point; (3) The hidden layer adopts a nonlinear activation function to extract nonlinear mapping characteristics of a high-order Bessel function; (4) The output layer adopts a linear activation function to directly output the predicted real-time modulation depth C.
  5. 5. The method for detecting a low-frequency vibration signal by using a neural network to assist triple mixing PGC differential demodulation according to claim 4, The logic of correcting the coefficient by using the real-time C value in S4 is as follows: ; Substitution recurrence formula ; Thus (2) And substituting the real-time modulation depth C into T, so that the influence of C value fluctuation can be eliminated.
  6. 6. The method of claim 5, wherein the signal interferometer and the reference interferometer are disposed in the same environment and are subject to the same environmental disturbance, i.e., common mode noise Δφ s ≈Δφ r, , thereby ; Since the signal interferometer and the reference interferometer are opposite in vibration sensitivity direction, phi s (t) = -φ r (t) exists, thereby ; The specific differential processing formula in S5 is: 。

Description

Neural network-assisted triple mixing PGC differential demodulation low-frequency vibration signal detection method Technical Field The invention belongs to the technical field of optical fiber sensor phase demodulation, and particularly relates to a neural network-assisted triple mixing PGC differential demodulation low-frequency vibration signal detection method. Background The interference type optical fiber sensor is widely applied to the fields of underwater sound detection, earthquake monitoring and the like. PGC (phase generated carrier) technology is its core demodulation means. However, existing demodulation schemes have some problems: 1. The accuracy of the conventional arctangent demodulation algorithm is highly dependent on the stability of the carrier modulation depth (C value), which is typically required to be stable around 2.63. In practical engineering applications, the actual C value tends to drift due to laser wavelength drift, ambient temperature changes, or drive circuit instability. When the C value deviates from the optimal operating point, the traditional double mixing PGC algorithm can generate serious nonlinear harmonic distortion, so that the signal-to-noise ratio is reduced, and even the signal cannot be correctly demodulated 2. Pure digital quadruple mixing algorithm, wherein the modulation depth C can be calculated in real time, but the calculation formula involves a large number of division operations. When the signal phase is at some specific point (e.g., the denominator approaches zero), severe periodic spikes in the computation result occur, and complex threshold filtering or heavy smoothing filtering must be relied upon, which reduces real-time and introduces additional noise. 3. The pure neural network method has a high processing speed and does not need mixing, but has limited generalization capability for complex waveforms, and linearity when the phase is widely changed is difficult to ensure. Disclosure of Invention Based on the problems in the prior art, the invention provides a neural network-assisted triple mixing PGC differential demodulation low-frequency vibration signal detection method. The advantages of the neural network and the triple mixing algorithm are combined, and a mixed architecture of the physical model driving feature extraction and the neural network performing parameter fitting is designed. The invention is realized by the following technical scheme: a neural network assisted triple mixing PGC differential demodulation low-frequency vibration signal detection method comprises the following steps: S1, acquiring interference signals output by two paths of opposite sensitive interference type optical fiber sensing systems, wherein the interference signals comprise a signal interferometer and a reference interferometer, the signal interferometer and the reference interferometer are arranged in the same closed environment and are opposite to the vibration sensitive direction, and the interference signals comprise Phase Generation Carrier (PGC) information determined by carrier modulation depth C; S2, extracting features, namely mixing two groups of interference signals with fundamental wave, second harmonic and third harmonic carrier signals respectively, and extracting orthogonal component feature vectors V= [ V 1,V2,V3 ] containing 1-3-order Bessel function items through low-pass filtering; s3, modulating depth fitting based on a neural network, namely, using the orthogonal component V extracted in the step S2 as input by the fully connected neural network FCNN, and directly outputting the real-time carrier modulating depth C at the current moment through a pre-trained nonlinear mapping relation; And S4, coefficient correction, namely calculating the quadrature component V 1、V2、V3 extracted in the S2 by utilizing the smooth carrier modulation depth C output in the S3 and combining with a Bessel function recurrence relation to construct a correction coefficient T for phase reduction, wherein the expression is as follows: ; V 1、V2、V3 is the quadrature components of the first harmonic, the second harmonic and the third harmonic extracted in the step S2; S5, differential restoration and unwrapping, namely carrying out triangular function domain differential processing on demodulation parameters of the signal interferometer and the reference interferometer before arc tangent, counteracting environmental interference by utilizing the characteristic that two paths of common mode noise are approximately equal, obtaining wrapped phases through arc tangent operation, and restoring continuous phase signals to be detected, which are not influenced by carrier modulation depth drift, through a phase unwrapping algorithm. Further, the expression of the light intensity signal I output by the signal interferometer and the reference interferometer in S1 is: ; Wherein A is direct current light intensity, B is alternating current term amplitude related to interference f