Dynamic demodulation method and system for MEMS optical fiber multi-parameter sensor

Abstract

The invention relates to the technical field of sensor demodulation, in particular to a dynamic demodulation method and system for an MEMS optical fiber multi-parameter sensor. The method comprises the following steps: deploying a high-speed parallel acquisition sensor for an MEMS optical fiber multi-parameter sensor, acquiring spectral reflection and spectral time domain data, and constructing an original multi-parameter spectral data set; performing signal separation processing on the original multi-parameter spectral data set based on a preset hybrid network to obtain a wavelength-light intensity matrix, an input layer based on the preset hybrid network being the original multi-parameter spectral data set, and an output layer being the wavelength-light intensity matrix; therefore, by constructing an integrated data flow of original spectrum data acquisition, feature decoupling, parameter fusion and performance evaluation, the problems of large coupling error, weak noise suppression and uncontrollable demodulation precision in a traditional optical fiber multi-parameter demodulation model are solved; and the demodulation accuracy and practicability of the MEMS optical fiber sensor under the multi-physical-quantity complex working condition are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sensor demodulation, and in particular to a dynamic demodulation method and system for a MEMS fiber multi-parameter sensor. BACKGROUND

[0002] The existing MEMS fiber multi-parameter sensor demodulation technology mainly relies on static modeling, separate decoupling and fixed parameter analysis method, and is widely used in fiber signal detection scenes of multiple physical quantities such as temperature, strain and pressure. However, this kind of technology has many key deficiencies, which limits the improvement of demodulation accuracy, response speed and system adaptability. First of all, the traditional demodulation method generally uses fixed frequency sampling and single frequency domain or time domain analysis method, which cannot effectively identify and track the non-stationary disturbance characteristics in the spectrum signal under complex dynamic environment, especially under high frequency disturbance or multi-source interference condition, the mutation and nonlinear response of the signal are easily ignored, causing distortion of feature extraction. Secondly, the existing technology mostly uses decoupling model with fixed structure, and lacks in-depth modeling ability for the coupling relationship between temperature, strain and pressure and other parameters, resulting in that the cross-sensitivity problem cannot be effectively suppressed, especially in the multi-field cooperative change scene, it is easy to cause measurement error accumulation and misdemodulation risk. In addition, some systems rely on static filtering and preset parameter control in data processing, lack of adaptive learning and response ability to statistical characteristics of input data, and are difficult to cope with the influence of noise background change and non-Gaussian disturbance, thereby affecting the robustness and precision consistency of demodulation results. More prominent is that the traditional fiber sensing demodulation system has low degree of software and hardware coupling, and cannot realize the lightweight and deployability of parameter model, resulting in poor real-time performance of the whole system, large resource overhead, and difficult to meet the engineering needs of edge deployment and fast response. SUMMARY

[0003] Therefore, it is necessary to provide a dynamic demodulation method and system for a MEMS fiber multi-parameter sensor to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a dynamic demodulation method for a MEMS fiber multi-parameter sensor, the method comprising the following steps:

[0005] Step S1: deploying a high-speed parallel acquisition sensor for a MEMS fiber multi-parameter sensor, acquiring spectrum reflection and spectrum time domain data, and constructing an original multi-parameter spectrum data set;

[0006] Step S2: based on the preset mixed network, the original multi-parameter spectrum data set is subjected to signal separation processing to obtain a wavelength-light intensity matrix, wherein the input layer of the preset mixed network is the original multi-parameter spectrum data set, and the output layer is the wavelength-light intensity matrix; the matrix of the decoupled spectrum component data is subjected to fiber lock-in amplification processing, and peak positioning of the reflection spectrum is performed to obtain decoupled fiber peak value data; the decoupled fiber peak value data is subjected to spectrum tuning to obtain the decoupled spectrum component data;

[0007] Step S3: the decoupled spectrum component data is subjected to dynamic time-frequency domain joint decoupling, and residual noise component elimination is performed to obtain a decoupled physical quantity characteristic matrix;

[0008] Step S4: model parameter fusion is performed based on the physical quantity characteristic matrix to obtain model parameter fusion data; error compensation is performed using the model parameter fusion data, and a MEMS fiber multi-parameter demodulation model is constructed;

[0009] Step S5: demodulation performance evaluation is performed based on the MEMS fiber multi-parameter demodulation model to obtain fiber sensor demodulation performance evaluation data; demodulation accuracy report generation is performed using the fiber sensor demodulation performance evaluation data.

[0010] The beneficial effects of the present application are that by deploying a high-speed parallel acquisition device, original signal data containing spectral reflection characteristics and time domain response characteristics are collected from the optical fiber, a high-resolution, multi-dimensional original spectrum data set covering temperature, strain and pressure and other multi-parameters is constructed; secondly, in the signal separation link, the spectrum components corresponding to each physical quantity are subjected to time-frequency domain coupling calculation through wavelet packet reconstruction, sparse representation and empirical mode decomposition, etc., which significantly improves the decoupling accuracy; on this basis, noise tracking and residual elimination are performed on the decoupled signal to construct a low-noise, reusable physical quantity characteristic matrix that can be used for subsequent modeling. In the modeling stage, a model parameter fusion strategy based on covariance weight adjustment and gradient regularization optimization is proposed to dynamically adjust the cross-sensitivity coefficients between each physical quantity to improve the physical consistency of feature expression and the convergence of data fitting. Subsequently, in the demodulation model construction and performance evaluation stage, an error compensation mechanism and a multi-dimensional data fusion algorithm are introduced to dynamically level the demodulation residual, and an error histogram, a response curve and a consistency scatter plot are constructed using the evaluation calibration data set to construct a demodulation accuracy evaluation system from a statistical level. Finally, all evaluation data are visualized and a standardized accuracy report is automatically generated to provide quantitative feedback basis for subsequent sensor deployment and model updating. Therefore, by constructing an integrated data flow of original spectrum data acquisition, feature decoupling, parameter fusion and performance evaluation, the present application solves the problems of large coupling error, weak noise suppression and uncontrollable demodulation accuracy in traditional fiber multi-parameter demodulation models, and improves the demodulation accuracy and practicality of MEMS fiber sensors in complex multi-physical quantity conditions.

[0011] Preferably, step S1 comprises the following steps:

[0012] Step S11: Deploy a multi-channel fiber Bragg grating array in the sensitive area of the MEMS fiber sensor, with a wavelength coverage range of 1520-1570 nm, a temperature range of -200℃-500℃, a strain range of 5000με, a pressure range of 0-3.5MPa, a sampling frequency of 8MHz, an accuracy of 0.1% FS, and collection of original optical spectrum reflection signals;

[0013] Step S12: Time-interleaved sampling of the original optical spectrum reflection signals using a photonic integrated ADC chip, with a single-channel sampling rate of 250kHz, a total sampling rate of 8MHz achieved by 32 channels in parallel, a quantization bit width of 18 bits, and generation of multi-channel spectral time-domain data;

[0014] Step S13: Baseline calibration of the multi-channel spectral time-domain data for environmental noise, and elimination of electromagnetic interference pulses through an adaptive threshold method, to generate denoised spectral time-domain data; spectral reconstruction of the denoised spectral time-domain data, conversion of the time-domain signals into wavelength-intensity distribution using a non-uniform Fourier transform, and generation of an original multi-parameter spectral data set.

[0015] The application can realize high-frequency (8MHz) real-time acquisition in the temperature range of -200 DEG C to 500 DEG C, the pressure range of 0-3.5MPa and the strain range of 5000με by deploying a multi-channel FBG array covering the wavelength range of 1520-1570nm in the sensor sensitive area, guaranteeing the integrity of the original optical spectrum reflection signal in the time and physical quantity dimensions. In the signal digitization process, 32-channel time-interleaved sampling is carried out by using an integrated photon ADC chip, on the basis of maintaining a single-channel sampling rate of 250kHz, a total sampling rate of 8MHz is realized through a parallel structure, combined with 18bit high-resolution quantization bit width, which effectively enhances the dynamic response recording ability of the low-intensity reflection peak, and improves the digital precision and fluctuation analysis ability of the data. Subsequently, in the noise processing link, an environmental noise baseline calibration mechanism is introduced to eliminate the background drift in the spectrum time domain with a dynamically updated reference baseline; at the same time, an adaptive threshold method is used to eliminate high-frequency interference pulses, so that abnormal peaks are accurately eliminated without losing effective signals, and stable denoised spectrum time domain data are obtained. In the data format conversion stage, nonlinear sampling correction is carried out based on the non-uniform Fourier transform (NUFFT), so that the time domain sampling points are accurately mapped to the wavelength-intensity distribution space, and the multi-channel reflection signal data with complete spectral characteristics are reconstructed, and the original multi-parameter spectrum data set for subsequent physical quantity decoupling and modeling is constructed. Therefore, by constructing a high-precision FBG array, a multi-channel interleaved sampling structure, a dynamic denoising and frequency domain conversion mechanism, the application solves the problems of insufficient frequency, quantization distortion and uncontrollable noise interference in the data acquisition of the traditional optical fiber sensor, and provides a high-fidelity, high-resolution original spectrum data basis for high-reliability demodulation model construction.

[0016] Preferably, the step S3 of dynamically decoupling the decoupled spectrum component data in the time and frequency domains comprises:

[0017] The adaptive noise covariance matrix is used to track signal mutations of the decoupled spectrum component data to obtain a spectrum component tracking signal, wherein the update frequency of the adaptive noise covariance matrix is 1kHz; the spectrum component tracking signal is processed in the time domain to obtain fiber sensor time domain processing data;

[0018] The decoupled spectrum component data is reconstructed by sparsification to obtain spectrum component reconstruction data; the spectrum component reconstruction data is processed in the frequency domain to obtain spectrum component frequency domain processing data.

[0019] The application realizes dynamic response enhancement of decoupled spectral components and multi-scale feature extraction by introducing an adaptive noise covariance matrix mechanism, combining sparse reconstruction and multi-domain data processing strategy, and effectively improves the precision and stability of the fiber sensor in sudden signal response, weak signal identification and frequency-time feature restoration. Firstly, aiming at the spectral shape nonlinear disturbance caused by temperature, strain or pressure mutation in the decoupled spectral component data, an adaptive noise covariance matrix is constructed and updated in real time, and the covariance estimation window is dynamically adjusted at a frequency of 1 kHz, which effectively suppresses the influence of environmental non-Gaussian noise, thereby tracking the tiny spectral line drift and burst signal with high precision, and generating a spectral component tracking signal. Subsequently, the tracking signal is subjected to sliding window integration, weighted filtering and mutation point enhancement operation in the time domain processing stage, and the fiber sensor response features with high time resolution are extracted, forming time domain processing data, and realizing accurate characterization of the time dimension change. At the same time, aiming at the redundancy and dimension expansion problem of the original decoupled data, the spectral components are compressed and sensed by the sparse reconstruction algorithm, which significantly improves the compactness and separability of the data expression and reduces the subsequent processing calculation amount. In the frequency domain processing, the spectral density estimation and bandwidth distribution analysis of the reconstructed data are carried out by using Fourier basis or wavelet basis, and the stable frequency domain principal component and weak frequency response feature are extracted, forming the spectral component frequency domain processing data, so as to realize the spectral structure analysis under the influence of multi-parameter change. In summary, the processing path combines multi-dimensional information tracking, adaptive denoising, sparse coding and frequency-time fusion analysis, and provides a multi-angle and high-precision data expression basis for subsequent modeling and demodulation.

[0020] Preferably, the residual noise component elimination in step S3 comprises:

[0021] Obtaining a noise dictionary library;

[0022] Matching and tracking the fiber sensor time domain processing data and the spectral component frequency domain processing data by using the noise dictionary library to obtain fiber noise tracking data;

[0023] Noise anomaly screening is performed on the fiber noise tracking data, and residual noise components are eliminated to obtain fiber noise screening data;

[0024] Signal regularity spectral density calculation is performed on the fiber noise screening data, the calculation standard is that SNR is greater than 35 dB, and a feature matrix is constructed to obtain a decoupled physical quantity feature matrix.

[0025] The application realizes a high-precision processing path of multi-domain joint noise recognition and residual noise suppression by constructing and calling a high-dimensional noise dictionary library, and gradually matching and tracking the time domain processing data and spectral component frequency domain processing data of the optical fiber sensor. Further, the application provides a data basis for constructing a stable and separable physical quantity feature matrix. Specifically, the noise dictionary library is constructed based on historical experimental sampling and multi-working condition simulation, and covers the frequency distribution mode and time evolution characteristics of multiple types of background noise (such as thermal noise, electromagnetic interference, light source jitter, scattering artifacts, etc.). In the noise tracking process, the dictionary library is compared with the current time domain processing data and frequency domain processing data in a dynamic sliding window, and algorithms such as cosine similarity and dynamic time warping (DTW) are used to identify and locate the noise mode, thereby forming fiber noise tracking data. Further, to enhance signal effectiveness, abnormal segments deviating from the statistical distribution of noise in the tracking data are multi-dimensionally screened, and residual noise is removed in combination with amplitude threshold discrimination and short-time energy mutation detection mechanisms to construct fiber noise screening data, thereby realizing efficient separation of superimposed and interpenetrated noise components. On this basis, to extract stable mapping characteristics between spectral response and physical disturbance, spectral density estimation is performed on the noise-screened data, power spectrum integration under multi-channel signal fusion is used, and low-quality spectral components are screened out by taking a signal-to-noise ratio (SNR) greater than 35 dB as a standard, and only signal segments with clear dominant frequency characteristics are retained for modeling. Finally, the feature vectors in the effective frequency spectrum interval are converted into a structured decoupled physical quantity feature matrix through statistical encoding and dimension normalization processing, thereby laying a data foundation for subsequent multi-parameter joint recognition and demodulation modeling.

[0026] Preferably, step S4 comprises the following steps:

[0027] Step S41: performing temperature-strain-pressure cross-sensitivity analysis on the physical quantity feature matrix, and performing model parameter fusion to obtain model parameter fusion data;

[0028] Step S42: performing reverse propagation fine-tuning processing on the model parameter fusion data to obtain reverse parameter fine-tuning data; performing decoupling accuracy judgment according to the reverse parameter fine-tuning data, and when the decoupling accuracy is less than 1.5%, performing 8-bit fixed-pointing on the reverse parameter fine-tuning data, and compressing the parameter quantity to less than 300,000 to obtain model lightweight parameter data;

[0029] Step S43: constructing a MEMS optical fiber multi-parameter demodulation model by using the model lightweight parameter data.

[0030] The application effectively identifies the collaborative disturbance and coupling interference mode between the parameters by constructing a high-resolution physical quantity feature matrix and performing temperature, strain and pressure cross-sensitivity analysis, and then fuses and reconstructs the high-correlation variables in the feature matrix to generate model parameter fusion data with low redundancy and high representation capability. Subsequently, a back propagation fine-tuning mechanism is introduced, and an iterative optimization method with gradient constraint is used to correct the fusion parameters by error back feedback, gradually approaching the optimal decoupling state, and taking the decoupling error threshold of 1.5% as the precision judgment standard. Under the condition of meeting the high decoupling accuracy, the parameter set after back propagation optimization is subjected to fixed-point processing, and the original high-bit-width floating-point parameters are compressed to 8-bit fixed-point format, and the parameter pruning strategy is applied to reduce the network redundancy structure, so as to control the overall parameter quantity to less than 300,000, effectively reducing the model calculation resource occupation and hardware deployment threshold. The finally generated model lightweight parameter data not only retains the high sensitive component response characteristics in the decoupling process, but also has good deployment adaptability and real-time running ability, providing a stable and efficient data basis for constructing a MEMS optical fiber multi-parameter demodulation model for edge side or low power consumption scene.

[0031] Preferably, step S41 comprises the following steps:

[0032] Step S411: performing normalization processing according to the physical quantity feature matrix to obtain a physical standardized feature matrix;

[0033] Step S412: performing temperature-strain-pressure covariance analysis and cross-sensitivity analysis using the physical standardized feature matrix to obtain an optical fiber cross-sensitivity coefficient matrix, wherein the non-diagonal elements of the optical fiber cross-sensitivity coefficient matrix represent temperature-strain data, temperature-pressure data and strain-pressure data;

[0034] Step S413: performing interference quantity extraction analysis on the temperature-strain data, temperature-pressure data and strain-pressure data to obtain optical fiber sensor interference quantity data;

[0035] Step S414: performing model parameter fusion on the optical fiber cross-sensitivity coefficient matrix using the optical fiber sensor interference quantity data according to the convergence condition of less than 0.5% and the iteration number of 10 to obtain model parameter fusion data.

[0036] The application constructs a physical standardization feature matrix by normalizing the physical quantity feature matrix, thereby ensuring that the three physical channels of temperature, strain and pressure have uniform dimensions and equal weights in numerical dimensions, and reducing the deviation interference caused by inconsistent data scales. Based on the standardized data matrix, a covariance analysis method is introduced to measure the mutual relationship between different physical quantities, and further extract three groups of non-diagonal line data of temperature-strain, temperature-pressure and strain-pressure as cross-sensitivity feature indexes to construct a fiber cross-sensitivity coefficient matrix. The matrix clearly represents the interaction interference mode and relative intensity in the multi-parameter sensing process, and provides accurate response basis for subsequent interference reduction. Then, the above three groups of coupled data are analyzed for independent interference quantity extraction, and the signal space projection and cooperative suppression method is used to separate the pure disturbance component from the interaction quantity, forming a recognizable interference quantity data set. In the fusion stage, an iterative optimization method based on interference quantity data is used to dynamically adjust the cross-sensitivity coefficient matrix through an error convergence control mechanism (error threshold less than 0.5%, and iteration times not more than 10), to realize directional suppression and mutual information minimization processing of the sensitive coupling component, and finally generate low correlation and high independence model parameter fusion data, which lays a high-quality input data foundation for subsequent decoupling model construction.

[0037] Preferably, the step S43 comprises the following steps:

[0038] Step S431: converting the model lightweight parameter data into floating-point number format to obtain FPGA executable parameter data;

[0039] Step S432: performing embedded logic unit mapping based on the FPGA executable parameter data to obtain optical fiber hardware model construction parameters;

[0040] Step S433: integrating into the optical fiber sensor to construct a demodulation model according to the optical fiber hardware model construction parameters, to obtain a MEMS optical fiber multi-parameter demodulation model.

[0041] The application converts model lightweight parameter data from fixed-point representation to floating-point format, generates FPGA executable parameter data, ensures the numerical stability and operation precision of data in hardware deployment, and provides a data format basis conforming to the FPGA operation system for subsequent logic unit mapping. The conversion process not only preserves the core parameter characteristics of the demodulation model, but also complies with the resource constraints of the embedded architecture through data format conversion, realizing smooth migration of the model from the training domain to the execution domain. Then, FPGA executable parameter data is used to construct resource mapping relationships for embedded logic units, and resource optimization configuration is performed among lookup tables (LUTs), digital signal processors (DSPs), and on-chip storage units (BRAMs), finally obtaining fiber hardware model construction parameters, thereby completing the process of mapping the demodulation model parameters to the hardware logic structure. This process ensures that the operation delay, power consumption and throughput of the demodulation model on the FPGA are within a controllable range, and realizes on-chip real-time processing of the spectrum-time domain data demodulation task. Finally, the fiber hardware model construction parameters are integrated into the fiber sensor system, realizing embedded deployment and functional integration of the demodulation model, enabling the sensor to have the ability to independently decouple and identify temperature, strain and pressure changes, and completing the complete closed-loop link from data acquisition, feature extraction to parameter demodulation. Therefore, through parameter lightweight conversion to floating point, FPGA mapping optimization and hardware integrated deployment, the application solves the problems of high dependence on computing power and poor embedded adaptability of traditional fiber demodulation models, and improves the real-time response capability and edge computing capability of the fiber sensor system.

[0042] Preferably, step S5 comprises the following steps:

[0043] Step S51: constructing an evaluation calibration data set based on the MEMS fiber multi-parameter demodulation model;

[0044] Step S52: performing demodulation performance evaluation using the evaluation calibration data set to obtain fiber sensor demodulation performance evaluation data;

[0045] Step S53: using the fiber sensor demodulation performance evaluation data to draw an error distribution histogram according to a bin number of 20 and a 95% confidence interval; using the fiber sensor demodulation performance evaluation data to generate a fiber response curve according to a time resolution of 1ms and superimposing the original multi-parameter spectrum data set; using the fiber sensor demodulation performance evaluation data to calculate demodulation values and construct a temperature-strain-pressure consistency scatter plot to obtain a fiber consistency scatter plot;

[0046] Step S54: visualizing the error distribution histogram, the fiber response curve and the fiber consistency scatter plot, and outputting a demodulation accuracy report.

[0047] The application verifies the performance of the MEMS fiber multi-parameter demodulation model by constructing an evaluation calibration dataset, first ensures that the evaluation data covers the multi-field coupling change state of three types of physical parameters of temperature, strain and pressure, and establishes a unified time sequence index and spectrum response label system, so that the evaluation calibration data has the characteristics of cross-parameter, multi-scale and high consistency. Then, the fiber sensor demodulation performance evaluation is carried out by using the dataset, the residual value between the model output and the true parameter is extracted, and the demodulation performance evaluation data is constructed, which can quantitatively evaluate the error stability and precision boundary of the demodulation model under different working conditions. On this basis, the evaluation data is statistically binned (bin number is 20) and combined with 95% confidence interval calculation to generate error distribution histogram, realizing the structured expression of error statistical law; at the same time, high-density fiber response curve is constructed with time resolution of 1ms, which is aligned with the original multi-parameter spectrum dataset to reveal the dynamic response characteristics and data fitting degree of the demodulation model; in addition, the demodulation value is mapped to the three-dimensional parameter space to form the temperature-strain-pressure consistency scatter diagram, which intuitively presents the spatial coordination degree and coupling relationship between the demodulation values through visualization. Finally, the above atlas results are summarized as a unified visualization index set to construct a demodulation accuracy report, which improves the readability of the demodulation result, the structural integrity of the evaluation result and the engineering applicability of the index comparison. Therefore, through the extraction, comparison and visualization of multi-level demodulation evaluation data, the application solves the problem of relying on manual interpretation and lacking of structured index system in traditional sensor demodulation performance evaluation method, and improves the performance explainability and application controllability of MEMS fiber sensor under complex working conditions.

[0048] Preferably, step S52 comprises the following steps:

[0049] Step S521: Calculate the temperature-strain-pressure demodulation value error of the evaluation calibration dataset, and construct the matrix distribution to obtain the fiber error distribution matrix;

[0050] Step S522: Record the response time test of the step signal of the evaluation calibration dataset to obtain the fiber response time test data;

[0051] Step S523: Extract the demodulation signal according to the evaluation calibration dataset; perform power spectrum analysis on the demodulation signal to obtain demodulation signal evaluation data, wherein the frequency band of the power spectrum analysis is 1-0 kHz;

[0052] Step S524: Randomly weight data fusion is performed on the fiber error distribution matrix, the fiber response time test data and the demodulation signal evaluation data to obtain the fiber sensor demodulation performance evaluation data.

[0053] The application improves the ability to describe the dynamic response and error behavior of the MEMS optical fiber sensor in a multi-parameter coupling environment by introducing a multi-dimensional demodulation performance evaluation method. First, according to the evaluation calibration data set, the errors between the demodulation values and the true values of temperature, strain and pressure are statistically calculated, and an optical fiber error distribution matrix is constructed to realize high-resolution mapping of errors in the parameter space, so that the error deviation trend of the model under the anisotropic condition in space can be clearly expressed. Secondly, by introducing a step signal to induce the sensor response process, the response time of the system under temperature, strain and pressure step disturbance is measured and recorded, and the optical fiber response time test data is generated, which reflects the tracking ability and dynamic lag characteristic of the demodulation model to the transient input, and helps to accurately characterize the behavior stability of the model under the mutation field. At the same time, the demodulation signal is further extracted from the evaluation calibration data set, and the power spectral density analysis is carried out in the frequency range of 1-1000 Hz to extract the frequency domain response characteristics of the system and obtain the frequency domain stability index of the demodulation signal, so as to realize the performance evaluation of the model in the noise suppression and frequency distribution dimension. Finally, the error distribution matrix, response time test data and power spectrum analysis results are integrated into a unified evaluation index set in a random weight fusion manner to generate a comprehensive demodulation performance evaluation data, which realizes the aggregation representation of performance characteristics while maintaining the independence of multi-dimensional data. The fusion data not only strengthens the lateral comparison of demodulation performance, but also facilitates subsequent quantitative evaluation and trend tracking between different working conditions and versions. Therefore, by constructing a fusion evaluation system of multi-source error, response and spectrum data, the application solves the technical bottleneck that the traditional demodulation performance evaluation is limited to single-dimensional error analysis and is difficult to express dynamic characteristics, and significantly improves the response ability description precision and multi-index comprehensive evaluation efficiency of the MEMS optical fiber sensor multi-parameter demodulation system in complex environment.

[0054] In the present specification, a dynamic demodulation system for a MEMS optical fiber multi-parameter sensor is provided for performing the above-mentioned dynamic demodulation method for a MEMS optical fiber multi-parameter sensor, and the dynamic demodulation system for a MEMS optical fiber multi-parameter sensor comprises:

[0055] An original data acquisition module is configured to deploy a high-speed parallel acquisition sensor for a MEMS optical fiber multi-parameter sensor, acquire spectral reflection and spectral time domain data, and construct an original multi-parameter spectral data set.

[0056] The signal decoupling and preprocessing module is configured to perform signal separation processing on the original multi-parameter spectrum data set based on a preset mixing network to obtain a wavelength-light intensity matrix, wherein an input layer of the preset mixing network is the original multi-parameter spectrum data set, and an output layer of the preset mixing network is the wavelength-light intensity matrix; perform fiber lock-in amplification processing on a matrix of the decoupled spectrum component data, and perform peak positioning on a reflection spectrum to obtain decoupled fiber peak data; and perform spectrum tuning on the decoupled fiber peak data to obtain the decoupled spectrum component data.

[0057] The dynamic time-frequency domain decoupling and feature extraction module is configured to perform dynamic time-frequency domain joint decoupling on the decoupled spectrum component data, and remove residual noise components to obtain a decoupled physical quantity feature matrix.

[0058] The model fusion and demodulation modeling module is configured to perform model parameter fusion based on the physical quantity feature matrix to obtain model parameter fusion data; perform error compensation using the model parameter fusion data, and construct a MEMS fiber multi-parameter demodulation model.

[0059] The performance evaluation and report generation module is configured to perform demodulation performance evaluation based on the MEMS fiber multi-parameter demodulation model to obtain fiber sensor demodulation performance evaluation data; and generate a demodulation accuracy report using the fiber sensor demodulation performance evaluation data.

[0060] The application realizes fine perception and intelligent interpretation of high-dimensional multi-parameter physical field by constructing a MEMS fiber multi-parameter demodulation system integrating original data acquisition, signal decoupling, feature extraction, parameter fusion modeling and performance evaluation. First, the original data construction module effectively captures spectral reflection and time domain signals under the support of high-frequency sampling and multi-channel parallel architecture, and constitutes an original multi-parameter spectral data set covering temperature, strain, pressure and other multi-physical quantity perception dimensions, laying a high-quality data foundation for subsequent decoupling analysis. Second, the signal decoupling and feature extraction module introduces a dynamic time-frequency domain joint analysis method to decompose the mixed signal under mixed interference into identifiable spectral components, and combines residual noise elimination and adaptive feature extraction strategy to construct a multi-channel, time-varying decoupled physical quantity feature matrix, thereby realizing explicit expression of weak correlation information between parameters. On this basis, the parameter fusion and demodulation modeling module reconstructs the high-dimensional mapping relationship between multi-parameters through model parameter fusion operation and error compensation mechanism, and under the premise of ensuring compression and deployability, constructs a MEMS fiber multi-parameter demodulation model with high decoupling accuracy and fast response speed. Finally, the performance evaluation and report generation module generates high-confidence demodulation performance evaluation results based on model output data through a systematic index evaluation process, and constructs a demodulation accuracy report in the form of error distribution, response curve, consistency scatter plot and other multi-dimensional visualization methods, realizing reliability verification and data traceability expression of the model in different physical field environments. Overall, the application connects the whole process channel from data acquisition, processing to evaluation from multi-parameter original data construction to high-dimensional performance index output, solves the technical bottlenecks of traditional fiber sensing demodulation systems in low data fusion degree, poor modeling closed loop and weak performance verification, and comprehensively improves the digitalization and intelligent processing level of MEMS fiber sensing system in complex application scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 It is a step flowchart of a dynamic demodulation method for a MEMS fiber multi-parameter sensor;

[0062] Figure 2 It is a step flowchart of a dynamic demodulation method for a MEMS fiber multi-parameter sensor; Figure 1

[0063] Figure 3 It is a schematic diagram of a multi-channel fiber Bragg grating array of a MEMS fiber sensor sensitive region;

[0064] Figure 4 It is a schematic diagram of a MEMS fiber sensor; Figure 3

[0065] The implementation, functional characteristics and advantages of the application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION​​

[0066] The technical method of the patent of the present application will be described clearly and completely below in conjunction with the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0067] In addition, the drawings are only schematic illustrations of the present application and are not necessarily drawn to scale. Identical reference signs in the drawings represent identical or similar parts, and thus repeated descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities, which do not necessarily have to correspond to physically or logically independent entities. The functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0068] It should be understood that although the terms "first", "second" and the like can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, without departing from the scope of the example embodiments, a first element can be called a second element, and similarly a second element can be called a first element. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0069] To achieve the above-mentioned purpose, please refer to Figures 1 to 4 A dynamic demodulation method for a MEMS optical fiber multi-parameter sensor, the method comprising the following steps:

[0070] Step S1: deploying a high-speed parallel acquisition sensor on the MEMS optical fiber multi-parameter sensor, acquiring spectral reflection and spectral time domain data, and constructing an original multi-parameter spectral data set;

[0071] Step S2: performing signal separation processing on the original multi-parameter spectral data set based on a preset mixing network, to obtain a wavelength-light intensity matrix, wherein the input layer of the preset mixing network is the original multi-parameter spectral data set, and the output layer is the wavelength-light intensity matrix; performing fiber lock-in amplification processing on the matrix of the decoupled spectral component data, and performing peak positioning on the reflection spectrum, to obtain decoupled fiber peak value data; performing spectral tuning on the decoupled fiber peak value data, to obtain decoupled spectral component data;

[0072] Step S3: performing dynamic time-frequency domain joint decoupling on the decoupled spectral component data, and performing residual noise component elimination, to obtain a decoupled physical quantity characteristic matrix;

[0073] Step S4: model parameter fusion based on the physical quantity feature matrix to obtain model parameter fusion data; error compensation is performed using the model parameter fusion data, and a MEMS fiber multi-parameter demodulation model is constructed;

[0074] Step S5: demodulation performance evaluation based on the MEMS fiber multi-parameter demodulation model to obtain fiber sensor demodulation performance evaluation data; demodulation accuracy report generation is performed using the fiber sensor demodulation performance evaluation data.

[0075] In the embodiment of the application, reference Figure 1 As shown in the figure, it is a step flow diagram of a dynamic demodulation method for a MEMS fiber multi-parameter sensor of the application. In this example, the dynamic demodulation method for the MEMS fiber multi-parameter sensor includes the following steps:

[0076] Step S1: deploying a high-speed parallel acquisition sensor for the MEMS fiber multi-parameter sensor, acquiring spectral reflection and spectral time domain data, and constructing an original multi-parameter spectral data set;

[0077] In the embodiment of the application, at the sensor level, the MEMS structure is coupled with the fiber transmission medium, and through multi-point deployment, the stress, temperature, pressure and other multiple physical quantities in the target environment are synchronously perceived, and the spectral response coding is performed through the reflective Bragg grating (FBG) or interference type structure. Secondly, the acquisition system adopts a high-bandwidth spectrum analyzer and a high-speed optical time domain reflectometer (OTDR) to construct an acquisition channel, wherein the spectral reflection data is sampled based on wavelength through a spectral analysis module, and the time domain data is time-resolved based on the pulse reflection path, and both are digitized and transmitted to the backend processing unit. In terms of parallel acquisition strategy, through the cooperation of multi-channel demultiplexing and high-speed analog-to-digital conversion (ADC) array, the synchronous acquisition of multiple sensor nodes is realized, and the time delay deviation and physical parameter coupling interference in data sampling are avoided. At the same time, in order to ensure the consistency and integrity of the data, a synchronization mechanism based on timestamp alignment is also introduced, the acquisition data of each channel is time-corrected, and the standardization data structure of spectral reflectivity and time domain waveform is constructed through the initial state reference value, and finally the original multi-parameter spectral data set with multi-source characteristics, time continuity and high-frequency dynamic change characteristics is formed, providing a unified data basis for subsequent signal decoupling and parameter modeling.

[0078] Step S2: signal separation processing is performed on the original multi-parameter spectrum data set based on a preset hybrid network to obtain a wavelength-light intensity matrix, wherein the input layer of the preset hybrid network is the original multi-parameter spectrum data set, and the output layer is the wavelength-light intensity matrix; fiber phase-locked amplification processing is performed on the decoupled spectrum component data matrix, and peak positioning of the reflection spectrum is performed to obtain decoupled fiber peak data; spectrum tuning is performed on the decoupled fiber peak data to obtain decoupled spectrum component data;

[0079] In the embodiment of the application, the original multi-parameter spectrum data set is sent into a hybrid network with a preset structure as an input, the network has a typical form of an input layer-hidden processing layer-output layer in structure, wherein the input layer is used to receive the original spectrum signal collected and containing the reflection characteristics of multiple physical quantities such as temperature, strain and pressure, the hidden layer realizes nonlinear deconstruction processing of the mixed signal by introducing signal mixing and decoupling weights and a feature mapping mechanism, so that the wavelength-light intensity matrix is obtained in the output layer, which is two-dimensionally reconstructed according to wavelength and light intensity. The matrix completely describes the reflection light intensity corresponding to each characteristic wavelength point at the data level, and is a key intermediate variable for extracting decoupled features. After obtaining the matrix, the decoupled spectrum component matrix is processed by applying the fiber phase-locked amplification technology. The phase-locked amplification technology realizes high-sensitivity enhancement of the target frequency component by phase synchronization of the reference signal and the to-be-measured signal, and significantly improves the signal-to-noise ratio of the spectrum signal by a noise suppression strategy. The data after the enhancement is subjected to peak positioning processing of the reflection spectrum, and the local maximum points in the spectrum line are identified by using the first derivative discrimination method or fitting interpolation and other means, so that the most representative spectrum reflection peak features are extracted to form the decoupled fiber peak data. On this basis, in order to ensure the physical consistency and spectral line stability in the subsequent demodulation modeling process, the decoupled fiber peak data is subjected to spectrum tuning processing. In this process, the original wavelength is adjusted according to the preset wavelength drift compensation function or fitting function, the peak shift caused by environmental disturbance or instrument error is corrected, and finally the decoupled spectrum component data which is stable and can be used for subsequent physical quantity mapping modeling is obtained. The processing flow converts the original data set into intermediate feature data with high signal-to-noise, clear physical characteristics and accurate spectral line positioning, which lays a key data foundation for subsequent decoupled modeling.

[0080] Step S3: dynamic time-frequency domain joint decoupling is performed on the decoupled spectrum component data, and residual noise components are removed to obtain a decoupled physical quantity feature matrix;

[0081] In the embodiment of the present application, at the time domain level, the processing flow segments the decoupled spectral component data for analysis based on window sliding technology or short-time Fourier transform (STFT), to obtain the instantaneous amplitude variation characteristics within each time window, reflecting the local dynamic behavior of spectral response at different time points; this operation allows identification of high-frequency disturbances, short-time mutations and other non-steady-state components. Subsequently, at the frequency domain level, the time domain signal is mapped to the frequency domain using fast Fourier transform (FFT) or wavelet transform (Wavelet Transform), to extract its frequency distribution structure, and further identify periodic components, harmonic components and inherent frequency response characteristics carried by each physical quantity. The two types of processing results are subsequently coupled in time and frequency through a feature mapping matrix or tensor structure, forming a joint expression at the data level. In order to eliminate residual noise components, the data after joint expression is further subjected to noise identification and suppression algorithms, such as threshold-based spectral subtraction, principal component analysis (PCA) noise reduction or independent component analysis (ICA) filtering processing, to maximize the separation and removal of background noise or coupled interference information unrelated to the target signal. Especially in the time-frequency coupling matrix, sub-band components with low energy density and unstable statistical characteristics are directly discarded or subjected to smoothing compression processing, to ensure that the final output data have stable physical feature expression capability. After the above processing is completed, the output data will be organized in matrix form, with each column representing the spectral characteristics at different time points, and each row corresponding to the time sequence variation of a specific wavelength or frequency component, thereby forming a decoupled physical quantity feature matrix, which serves as the core data basis for subsequent physical parameter mapping and modeling analysis.

[0082] Step S4: model parameter fusion based on the physical quantity feature matrix to obtain model parameter fusion data; error compensation using the model parameter fusion data, and construction of a MEMS fiber multi-parameter demodulation model;

[0083] In the embodiment of the present application, the decoupled physical quantity characteristic matrix obtained from step S2 is taken as input, and a multi-parameter parameter fusion framework is constructed at the data level to realize the close coupling and information complementation of the response characteristics of various physical fields in a unified space. Specifically, for the feature vectors corresponding to different physical quantities such as temperature, stress and pressure, dimension reduction and covariance structure modeling of the feature matrix are performed through multivariate statistical methods such as principal component analysis (PCA) or canonical correlation analysis (CCA) to extract the collaborative change pattern between different physical fields; at the same time, a weighted fusion strategy is used to input the discriminability of each component under different experimental conditions as the weight to construct a weighted average or a fusion operator based on Bayesian estimation, map the multi-source features to a unified fusion data space, and generate model parameter fusion data with the characteristics of minimized information redundancy and maximized discriminative information. Subsequently, in order to correct system errors and instrument calibration bias, an error compensation model needs to be constructed based on the fusion data. Common methods include linear regression compensation based on least squares, nonlinear compensation based on support vector regression (SVR) or Gaussian process regression (GPR), and online iterative error correction combined with Kalman filtering or extended Kalman filtering (EKF). By modeling and predicting the residual sequence between the fusion data and the calibration standard value, dynamic correction of systematic errors is realized. Finally, after error compensation, the corrected high-consistency, multi-dimensional fusion data is input into the MEMS fiber multi-parameter demodulation model. The model can use a semi-analytical model based on physical mechanism (such as embedding the demodulation formula into the finite element simulation result) or a data-driven deep learning model (such as a multi-input branch convolutional neural network or a long short-term memory network). By feature mapping and pattern recognition of the fusion data, the demodulation results of each physical quantity are obtained, thereby constructing an integrated and robust MEMS fiber multi-parameter demodulation model.

[0084] Step S5: Perform demodulation performance evaluation based on the MEMS fiber multi-parameter demodulation model to obtain fiber sensor demodulation performance evaluation data; and use the fiber sensor demodulation performance evaluation data to construct a demodulation accuracy report.

[0085] In the embodiment of the present application, the performance evaluation process first compares the output results of the MEMS fiber multi-parameter demodulation model with the reference data set collected in advance and strictly calibrated point by point. The corresponding data pairs are indexed by time stamp and spatial position, and the demodulation error sequence is generated by the residual calculation module; then, a statistical analysis process is constructed based on the error sequence, including root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R 2) and other scalar indicators are calculated to quantify the demodulation deviation and fitting degree of the model under the whole test condition, and the skewness and kurtosis of the error sequence are analyzed to reveal the probability distribution characteristics, and the stability and confidence interval of the system response are revealed by drawing the error probability density function (PDF) and cumulative distribution function (CDF) atlas; on this basis, in order to identify the performance difference of the model under different physical quantity range and dynamic change rate, the error data needs to be grouped and counted according to the preset partition (such as temperature gradient interval or stress amplitude level), and the dispersion degree and abnormal value state of the error distribution in each partition are observed through box plot or small batch statistical analysis; in addition, combined with the time series comparison analysis under the typical working condition, the frequency domain response indicators such as frequency response function (FRF) fitting degree evaluation can be introduced to judge the tracking ability of the model to the fast dynamic change signal; finally, the above evaluation results, including scalar indicators, partition statistical table, distribution atlas and frequency domain analysis data, are summarized as structured evaluation data set, and the bottom layer data file of demodulation accuracy report is generated according to the pre-defined report template field (such as indicator name, calculation method, result value, corresponding atlas link, etc.); the report generation module relies on the data file to embed scalar and graphic elements into the standard document structure through automatic script, forming a complete demodulation accuracy report including abstract table, performance comparison graph, error distribution graph and conclusion explanation section, providing comprehensive and traceable data support for subsequent experimental verification and system optimization.

[0086] Preferably, step S1 comprises the following steps:

[0087] Step S11: deploying a multi-channel fiber Bragg grating array in the sensitive area of the MEMS fiber sensor, with a wavelength coverage range of 1520-1570 nm, a temperature range of -200℃ to 500℃, a strain range of 5000με, a pressure range of 0-3.5MPa, a sampling frequency of 8MHz, an accuracy of 0.1%FS, and collecting original spectrum reflection signals;

[0088] Step S12: using a photonic integrated ADC chip to perform time-interleaved sampling on the original spectrum reflection signals, with a single-channel sampling rate of 250kHz, a total sampling rate of 8MHz realized by 32 channels in parallel, a quantization bit width of 18bits, and generating multi-channel spectrum time domain data;

[0089] Step S13: performing environmental noise baseline calibration on the multi-channel spectrum time domain data, and removing electromagnetic interference pulses through an adaptive threshold method to generate denoised spectrum time domain data; performing spectrum reconstruction on the denoised spectrum time domain data, converting the time domain signal into wavelength-intensity distribution by using non-uniform Fourier transform, and generating an original multi-parameter spectrum data set.

[0090] In the embodiment of the application, a multi-channel Bragg grating (FBG) array is arranged along the sensing structure in the sensitive area of the MEMS optical fiber sensor, the center wavelength of each FBG is distributed in the range of 1520 nm to 1570 nm, and the FBG corresponds to three physical quantity dimensions of temperature sensing (-200℃ to 500℃), strain sensing (0 to 5000με) and pressure sensing (0 to 3.5MPa) respectively; the reflection spectrum signal output by the array is collected as original spectral reflection data at a sampling rate of 8MHz and a quantization accuracy of 0.1%FS to form an initial spectral time-domain data matrix of NXM (N is the number of time sampling points, and M is the number of channels). Subsequently, the original matrix is time-interleaved sampled by an integrated photon ADC chip in step S12: each channel is independently sampled at 250kHz, and after 32 parallel channels, a multi-channel spectral time-domain data set with a total sampling rate of 8MHz is synthesized in the digital domain with a bit width of 18 bits; in this process, clock jitter and inter-channel time delay are first corrected by an internal delay-locked loop (DLL), and the relative time stamps of each channel are recorded in the digital domain to construct a synchronized multi-channel data stream. Next, the environmental noise baseline is calibrated at the data level in step S13: the static reference frame at the beginning of the collection is used to calculate the channel-by-channel baseline vector, the time variation trend is estimated by sliding window mean filtering, and the burst electromagnetic interference pulse is marked and removed by using the adaptive threshold method based on the noise power spectral density (PSD), and the denoised spectral time-domain data matrix is output; on this basis, for each calibrated time-domain signal, the non-uniform fast Fourier transform (NUFFT) algorithm is called to process the non-linear time interval in the FBG reflection pulse, the corrected time-domain sampling points are directly mapped to the non-equidistant wavelength coordinates, and the two-dimensional distribution data corresponding to the wavelength-intensity are generated, and finally the original multi-parameter spectral data set with the characteristics of multi-physical quantity response, time continuity and high resolution is formed, which provides a unified and structured data input for subsequent signal decoupling and feature extraction.

[0091] Preferably, the dynamic time-frequency domain joint decoupling of the decoupled spectral component data in step S3 comprises:

[0092] The adaptive noise covariance matrix is used to track the signal mutation of the decoupled spectral component data to obtain a spectral component tracking signal, wherein the update frequency of the adaptive noise covariance matrix is 1kHz; the spectral component tracking signal is processed in the time domain to obtain fiber sensor time domain processing data;

[0093] The decoupled spectral component data is sparsely reconstructed to obtain spectral component reconstruction data; the spectral component reconstruction data is processed in the frequency domain to obtain spectral component frequency domain processing data.

[0094] In the embodiment of the present application, based on the extended Kalman filter (EKF) or adaptive Kalman filter (AKF) framework, the spectral component signal is regarded as a state vector, and the observation noise covariance matrix is dynamically updated according to the residual sequence and innovation sequence at the previous moment, and the update frequency is set to 1 kHz to ensure high time resolution response to sudden spectral drift or reflection peak jump. In each iteration, through Kalman gain calculation, rapid gain adjustment of the instantaneous mutation signal is realized, and a smooth spectral component tracking signal is output; the signal then enters the time domain processing module, mainly including bandpass / lowpass filtering based on finite impulse response (FIR) or infinite impulse response (IIR) digital filter, time domain envelope detection, and sliding window mean or median detrending algorithm, to remove high-frequency noise residues and baseline drift, and output structured optical fiber sensor time domain processing data. At the same time, in order to recover the sparse spectral component structure information, the system implements sparse reconstruction on the same decoupled data set, usually using the compressed sensing (CS) method based on norm minimization or the sparse Bayesian learning (SBL) framework, by constructing a dictionary matrix DDD (such as a multi-scale spectral atom library obtained by statistical training data) and solving an optimization problem. The sparse coefficient vector is obtained, and the spectral component reconstruction data is reconstructed. Finally, the reconstruction data is processed in the frequency domain, mainly relying on fast Fourier transform (FFT) or high-resolution spectrum estimation algorithm (such as MUSIC, Capon method) to map the time domain signal to the frequency domain, and combining power spectral density (PSD) estimation and amplitude / phase unwrapping technology, the amplitude and phase changes of each frequency component are quantitatively extracted, and the spectral component frequency domain processing data is finally output to support subsequent physical quantity demodulation and feature analysis. This whole process realizes closed-loop processing from non-stationary signal mutation tracking to sparse structure recovery and frequency domain feature extraction at the data level, ensuring the time-frequency domain integrity and distinguishability of spectral component information.

[0095] Preferably, the residual noise component elimination in step S3 comprises:

[0096] Obtaining a noise dictionary library;

[0097] Matching and tracking the optical fiber sensor time domain processing data and the spectral component frequency domain processing data using the noise dictionary library to obtain optical fiber noise tracking data;

[0098] Performing noise anomaly screening on the optical fiber noise tracking data and eliminating residual noise components to obtain optical fiber noise screening data;

[0099] Performing signal regularity spectrum density calculation on the optical fiber noise screening data, calculating the standard as SNR greater than 35 dB, and constructing a feature matrix to obtain a decoupled physical quantity feature matrix.

[0100] In the embodiment of the present application, a noise dictionary library containing various typical noise patterns is constructed by offline or online methods. The library is obtained by preprocessing noise signal segments collected from experimental environments and actual working conditions, extracting multi-scale and multi-resolution noise primitives by wavelet packet decomposition or short-time Fourier transform, and training and updating the dictionary using the K-SVD algorithm or sparse auto-encoding network to ensure the coverage and adaptability of the dictionary to the noise characteristics of the target system. Then, the time domain processing data and spectral component frequency domain processing data of the optical fiber sensor are used as the signals to be matched, and the dictionary matching algorithm based on orthogonal matching pursuit (OMP) or sparse representation classification (SRC) is performed on the noise dictionary through the sliding window mechanism to calculate the projection coefficients of the signals on each noise primitive in real time, thereby obtaining optical fiber noise tracking data reflecting the dynamic distribution of the similarity between the signals and the noise patterns in the time-frequency domain over time or frequency. Next, an anomaly detection strategy is applied to the tracked noise component sequence, mainly based on the local outlier factor (LOF) or the statistical threshold method based on kurtosis and skewness, to screen and mark the sudden increase or distribution deviation of the projection coefficients, and to remove the corresponding residual noise components. The optical fiber noise screening data is generated by reconstructing the time domain / frequency domain signals or removing the noise primitive mapping at the sparse coefficient level. Finally, the power spectral density (PSD) of the screened signal sequence is estimated, usually using the Welch method or multiple spectrum estimation techniques, to calculate the signal-to-noise ratio (SNR) of each component signal, and only the frequency bands or time period segments that satisfy SNR≥35dB are retained. Based on this high signal-to-noise area, statistical quantities (such as time domain envelope energy, time-frequency center frequency deviation, spectral peak amplitude, etc.) and sparse representation coefficients are extracted in both time domain and frequency domain representation dimensions, and the resulting feature vectors are arranged according to physical quantity categories, time windows, or frequency band indices to construct the final decoupled physical quantity feature matrix, providing a clear and high-resolution multi-source data basis for subsequent model parameter fusion and error compensation.

[0101] As an example of the present application, reference is made to Fig. 1, which shows a schematic diagram of a method for decoupling physical quantities according to an embodiment of the present application. In this example, the step S4 includes: Figure 2

[0102] Step S41: Perform temperature-strain-pressure cross-sensitivity analysis on the physical quantity feature matrix, and perform model parameter fusion to obtain model parameter fusion data;

[0103] Step S42: Perform backpropagation fine-tuning processing on the model parameter fusion data to obtain backpropagation parameter fine-tuning data; perform decoupling accuracy judgment according to the backpropagation parameter fine-tuning data, when the decoupling accuracy is less than 1.5%, perform 8-bit fixed-pointing on the backpropagation parameter fine-tuning data, and compress the parameter quantity to less than 300,000 to obtain model lightweight parameter data;

[0104] ​Step S43: constructing a MEMS fiber multi-parameter demodulation model by using the model lightened parameter data.

[0105] In the embodiment of the present application, the collaborative variation pattern between each physical quantity is revealed by tensor decomposition or high-order principal component analysis (HOPCA), and a cross-sensitivity matrix is constructed based on multiple linear regression (MLR) or canonical correlation analysis (CCA) to quantify the response influence of each physical quantity on other measured signals. Then, the weight coefficients in the sensitivity matrix are mapped with the principal component scores in the feature matrix by using a weighted fusion strategy, and a unified model parameter fusion data is generated by matrix multiplication or tensor contraction, which retains the discriminative features of each physical field and eliminates the redundant coupling. The model parameter fusion data is input into a fine-tuning process based on back propagation: a loss function (such as root mean square error RMSE or mean absolute percentage error MAPE) composed of the fusion data and the standard calibration value drives the neural network parameter update, and the back parameter fine-tuning data is iteratively calculated by mini-batch stochastic gradient descent (mini-batch SGD) or adaptive optimizer (such as Adam), and the current decoupling accuracy is evaluated after each iteration. When the evaluation result shows that the decoupling accuracy is lower than the threshold of 1.5%, the model lightening process is triggered: first, the floating-point parameters are mapped to 8-bit fixed-point format under the static quantization framework, and the quantization error is minimized by using symmetric quantization or asymmetric quantization strategy combined with zero-point calibration; then, the structured pruning or unstructured pruning method is applied, and the redundant neurons and channels are removed by importance score (such as sensitivity analysis based on weight gradient) until the total amount of model parameters is compressed to less than 300,000, to generate the final model lightened parameter data. Finally, the quantized, pruned and quantized lightened parameters are loaded into the MEMS fiber multi-parameter demodulation model architecture, which efficiently performs forward inference through integer arithmetic units and fixed-point activation functions to realize efficient computation and deployment of the demodulation process, and lays a data level foundation for real-time multi-parameter sensing in resource-constrained environments.

[0106] Preferably, step S41 comprises the following steps:

[0107] Step S411: normalizing according to the physical quantity feature matrix to obtain a physical standardized feature matrix;

[0108] Step S412: performing temperature-strain-pressure covariance analysis and cross-sensitivity analysis by using the physical standardized feature matrix to obtain a fiber cross-sensitivity coefficient matrix, wherein the non-diagonal elements of the fiber cross-sensitivity coefficient matrix represent temperature-strain data, temperature-pressure data and strain-pressure data.

[0109] Step S413: Interference quantity extraction analysis is performed on the temperature-strain data, temperature-pressure data and strain-pressure data to obtain fiber sensor interference quantity data.

[0110] Step S414: Model parameter fusion is performed on the fiber sensor interference quantity data to obtain model parameter fusion data, according to a convergence condition less than 0.5% and an iteration number of 10.

[0111] In the embodiment of the present application, the decoupled physical quantity characteristic matrix is processed column by column to normalize, the original characteristic values of each channel are subtracted by the mean value and divided by the standard deviation, and a physical standardized characteristic matrix with zero mean and unit variance is constructed to eliminate the influence of the dimensional difference of each physical quantity on subsequent statistical analysis. Then, the covariance matrix between the three groups of characteristics of temperature, strain and pressure is calculated based on the standardized matrix, the linear correlation degree between each pair of characteristics is evaluated by the covariance value, and the non-diagonal elements corresponding to the temperature-strain, temperature-pressure and strain-pressure pairs are interpreted in detail by using canonical correlation analysis (CCA) or high-order cross spectrum analysis to form a fiber cross-sensitivity coefficient matrix. The three pairs of cross-sensitivity data are input into an interference quantity extraction module, and the error component separation is performed on the non-diagonal line items of the covariance matrix by using multivariate regression residual analysis or interference elimination method based on principal component tracking to extract interference quantity data reflecting the cross interference between each physical quantity. Finally, the obtained interference quantity data and the original standardized characteristic vector are input into the iteration optimization process of model parameter fusion, which takes the minimum absolute value of the non-diagonal line of the cross-sensitivity coefficient matrix to the convergence threshold value of 0.5% as the target, and uses weighted least squares method (WLS) or constraint optimization algorithm based on Lagrange multiplier method for iteration, and updates the fusion weight each time until the iteration number limit of 10 or the convergence criterion is met, and generates the final model parameter fusion data by tensor contraction or matrix weighted summation, so as to complete the quantitative suppression and fusion optimization of cross interference at the data level.

[0112] Preferably, step S43 comprises the following steps:

[0113] Step S431: converting the model lightweight parameter data into floating point number format to obtain FPGA executable parameter data;

[0114] Step S432: performing embedded logic unit mapping based on the FPGA executable parameter data to obtain fiber hardware model construction parameters;

[0115] Step S433: integrating into the fiber sensor to construct the demodulation model according to the fiber hardware model construction parameters to obtain the MEMS fiber multi-parameter demodulation model.

[0116] In the embodiment of the present application, the quantized and pruned lightweight model parameter data is format-converted to meet the interface specifications of the internal floating-point operation unit or fixed-point operation unit of the FPGA: specifically, the 8-bit fixed-point weights and activation parameters in the model are mapped back to numerical expressions conforming to the IEEE 754 or custom floating-point format through a dequantization algorithm, and the exponent bit and mantissa bit width are reconfigured and aligned according to the requirements of the hardware description library (IP Core) provided by the FPGA manufacturer, to generate FPGA executable parameter data that can be recognized by a high-level synthesis (HLS) tool; then, with the help of the HLS tool or a synthesis process based on a hardware description language (such as VHDL / Verilog), the FPGA executable parameter data is mapped to embedded logic units (LUTs), digital signal processor (DSP) blocks and block RAM (BRAM) resources: first, a computation graph description file is generated according to the dimensions and channel parallelism of each layer weight matrix, and resource allocation algorithm is used to automatically allocate computation core positions under timing constraints and power budget, and operations such as multi-input convolution or matrix multiplication are split into several data flow modules that can be executed in parallel; then, through the logic synthesis and layout routing process, the parameter storage address, bus width and clock domain crossing are finely configured, and the fiber hardware model construction parameters are output, including LUT configuration bitstream, DSP connection topology and BRAM data initial loading file; finally, the fiber hardware model construction parameters are integrated into the FPGA device of the MEMS optical fiber sensor, and through the software interface or board-level description file (such as Xilinx Vivado's.bit and.hwh files) of the embedded system, the hardware deployment of the demodulation model is realized: after loading the bitstream and starting the clock in the sensor control board, the FPGA inside can follow the preset pipeline structure and parallel computing strategy to receive the preprocessed spectral feature data in real time, perform model forward inference and output temperature, strain and pressure quantization results, thereby completing the whole process of hardware construction of the MEMS optical fiber multi-parameter demodulation model.

[0117] Preferably, step S5 comprises the following steps:

[0118] Step S51: Constructing an evaluation calibration data set based on the MEMS optical fiber multi-parameter demodulation model;

[0119] Step S52: performing demodulation performance evaluation using the evaluation calibration data set to obtain optical fiber sensor demodulation performance evaluation data;

[0120] Step S53: draw an error distribution histogram with 20 bins and 95% confidence interval; generate a fiber response curve by resampling the demodulation performance evaluation data with 1ms time resolution and superimposing the original multi-parameter spectrum data set; calculate the demodulation values and construct the temperature-strain-pressure consistency scatter plot to obtain the fiber consistency scatter plot;

[0121] Step S54: visualize the error distribution histogram, fiber response curve and fiber consistency scatter plot, and output the demodulation accuracy report.

[0122] In the embodiment of the application, for the calibration and evaluation requirements of the MEMS fiber multi-parameter demodulation model, the original multi-parameter spectrum data collected from the laboratory or the field and the corresponding temperature, strain and pressure standard reference values are used to construct the evaluation calibration data set: the time stamp, collected spectrum, model demodulation output and corresponding calibration instrument reading of each test are stored in a structured table, and time alignment and missing value interpolation are completed in the data preprocessing stage to ensure one-to-one correspondence and integrity of each entry in the evaluation data set; then, the demodulation model is evaluated by using the data set, all samples are traversed by programming script, the demodulation error sequence is calculated point by point, and the fiber sensor demodulation performance evaluation data including RMSE, MAE, mean deviation and standard deviation are obtained; on this basis, the error values are divided into equal width intervals according to the number of bins 20, the number of samples in each interval is counted, and the 95% confidence interval of each error box is estimated based on the Poisson distribution assumption or Bootstrap resampling method, so as to draw the error distribution histogram with confidence interval band; at the same time, the evaluation data is resampled according to the 1ms time window, and after alignment with the original multi-parameter spectrum data set, it is superimposed and displayed in the same chart to generate a high time resolution fiber response curve and reveal the tracking ability of the model to fast dynamic changes; in addition, by pairing the model demodulation values and the standard reference values in the temperature, strain and pressure three-dimensional, a consistency scatter plot is drawn to visualize the linear fitting relationship and dispersion characteristics of each physical quantity; finally, the error distribution histogram, fiber response curve and consistency scatter plot are typeset according to the pre-defined template by calling the visualization script or report automation tool, and the key scalar indicators (such as RMSE, R 2 value and confidence interval range) are embedded in the report as a table or annotation, and a complete demodulation accuracy report document containing chart title, coordinate label, legend and brief data description is generated, realizing the whole process closed loop from data collection to visualization analysis to report generation.

[0123] Especially important is that step S51 includes the following:

[0124] Obtain the design parameters of the MEMS fiber sensor;

[0125] The MEMS fiber sensor design parameters are divided according to temperature, strain and pressure intervals, and extreme value injection is performed to obtain a calibration working condition parameter table;

[0126] The simulation processing is performed on the calibration working condition parameter table, and labeling is added to obtain an evaluation calibration data set.

[0127] In the embodiment of the application, the design parameters of the MEMS fiber sensor are obtained in the form of a structured table from the sensor design document and the experimental calibration report, including the fiber Bragg grating (FBG) center wavelength range, grating length, refractive index sensitivity coefficient, base material thermal expansion coefficient and mechanical elastic modulus, and the corresponding temperature, strain and pressure nominal range and resolution. Subsequently, these design parameters are divided according to the equidistant or adaptive interval strategy along the three dimensions of temperature (such as -200℃ to 500℃), strain (0-5000με) and pressure (0-3.5MPa) respectively at the data level, and extreme value samples (including upper limit, lower limit and small perturbation beyond the range) are injected at the boundary and inside of each cell, thereby generating a multi-dimensional calibration working condition parameter table containing normal calibration points and boundary limit scenarios; the table is stored in the form of an N×3 matrix structure, each row represents a three-tuple working condition parameter, and is attached with parameter number and injection type label. Then, through a batch simulation process, digital twin simulation is performed on each row of data in the calibration working condition parameter table based on finite element analysis (FEA) or optical interference simulation software (such as the optical-mechanical coupling module in COMSOL Multiphysics), the corresponding temperature field, strain distribution and pressure load conditions are input, the Bragg wavelength drift and reflection spectrum change inside the sensor are simulated, and the simulation output includes time domain reflection waveform, spectral response curve and local stress-temperature coupling field graph. The simulation results are associated with the original parameter table in the data pipeline in JSON or HDF5 format, and automatically attached with true value labels (such as “temperature true value”, “strain true value”, “pressure true value”) and simulation output labels (such as “simulated peak wavelength”, “reflection intensity change”), and finally form a complete evaluation calibration data set. The whole process is managed through scripting at the data level, including parameter table generation script, simulation calling interface and labeling processing script, to ensure that each calibration working condition is traceable and has complete input-output mapping relationship, providing a strict data basis for error calculation and model calibration of subsequent performance evaluation.

[0128] Especially important is that the simulation processing of the calibration working condition parameter table includes the following:

[0129] The FBG wavelength shift of the calibration working condition parameter table is calculated to obtain an ideal spectral reflection peak;

[0130] The actual noise processing is performed by using an ideal spectral reflection peak with an environmental noise of -50dBm / Hz, and a noisy spectral signal is obtained;

[0131] The noisy spectral signal is simulated by a photon ADC, and spectral-time domain data conversion is performed to obtain a simulation calibration data set.

[0132] In the embodiment of the application, the corresponding center wavelength drift is calculated based on each row of parameters (temperature, strain, pressure triplets) according to the Bragg grating theory, so as to obtain the ideal spectral reflection peak position and amplitude distribution; then, the ideal reflection peak signal is superimposed with a white noise model in the frequency domain, the noise power spectral density is set to -50dBm / Hz, a random noise sequence conforming to the Gaussian distribution is generated and weighted according to the sampling bandwidth to obtain a noisy spectral signal containing environmental noise interference; then, in the photon integrated ADC simulation link, the noisy signal is subjected to analog front-end processing: first, discrete time sampling points are generated according to the preset sampling rate and clock jitter model, then the sampling values are truncated and rounded by applying the quantization bit width (such as 18bit) and dynamic range limit to simulate the quantization error of the actual ADC, and the time domain impulse response of each sampling is recorded; finally, the quantized spectral data is converted into time domain format by the non-uniform time-wavelength mapping algorithm, usually the wavelength coordinates are mapped back to the isochronous domain sampling points by using the inverse NUFFT or interpolation algorithm to form a complete simulation calibration data set. The entire process is sequentially responsible for wavelength shift calculation, noise superposition and ADC simulation by a scriptable module at the data level, and the ideal reflection peak, noise model parameters, quantization residual and time domain reconstruction results are stored together through a unified data structure (such as an HDF5 file), providing a repeatable and traceable simulation data basis for downstream performance evaluation and model training.

[0133] Preferably, step S52 comprises the following steps:

[0134] Step S521: Calculate the demodulation value error of temperature-strain-pressure according to the evaluation calibration data set, and construct a matrix distribution to obtain a fiber error distribution matrix;

[0135] Step S522: Record the response time test of the evaluation calibration data set to obtain fiber response time test data;

[0136] Step S523: Extract the demodulation signal according to the evaluation calibration data set; perform power spectrum analysis on the demodulation signal to obtain demodulation signal evaluation data, wherein the frequency band of the power spectrum analysis is 1-0kHz;

[0137] Step S524: Random weight data fusion is performed on the fiber error distribution matrix, the fiber response time test data and the demodulation signal evaluation data to obtain fiber sensor demodulation performance evaluation data.

[0138] In the embodiment of the present application, for each sampling point in the evaluation calibration dataset, the model demodulation output value is paired with the corresponding standard true value in the three dimensions of temperature, strain and pressure, all error values are filled into a three-dimensional matrix structure by calculating the error and layering according to the physical quantity and error size, and a fiber error distribution matrix is formed; then, a step input signal (temperature, strain or pressure mutation) is applied to the sensing system according to the evaluation calibration dataset, and the demodulation model response time is recorded, the threshold method or the percentage rise time (such as 10%-90% jump time) is used to measure and summarize the response delay of each step test section, and a fiber response time test data vector is generated; then, a continuous demodulation signal sequence is extracted from the evaluation calibration dataset and power spectrum analysis is performed thereon, usually the Welch method is used to segment and overlap the signal with a window function, fast Fourier transform (FFT) calculation is performed, and the power spectral density (PSD) is estimated within the frequency band of 1Hz to 10kHz, thereby obtaining a demodulation signal evaluation dataset-a two-dimensional array containing frequency components and their corresponding power values; finally, in order to obtain comprehensive performance evaluation results, the fiber error distribution matrix, the response time test data and the power spectrum evaluation data are weighted and fused according to a preset random weight vector: first, the three types of data are normalized to the same dimension (for example, maximum-minimum normalization or Z-score standardization), then weighted summation is performed at the element level or matrix level, and finally joint fiber sensor demodulation performance evaluation data reflecting error distribution, dynamic response and frequency domain characteristics are obtained for subsequent report generation and system optimization.

[0139] In the present specification, a dynamic demodulation system for a MEMS fiber multi-parameter sensor is provided for performing the dynamic demodulation method for a MEMS fiber multi-parameter sensor described above, and the dynamic demodulation system for a MEMS fiber multi-parameter sensor comprises:

[0140] An original data acquisition module is configured to deploy a high-speed parallel acquisition sensor to the MEMS fiber multi-parameter sensor, acquire spectral reflection and spectral time domain data, and construct an original multi-parameter spectral dataset.

[0141] A signal decoupling and preprocessing module is configured to perform signal separation processing on the original multi-parameter spectral dataset based on a preset mixing network to obtain a wavelength-intensity matrix, wherein the input layer of the preset mixing network is the original multi-parameter spectral dataset, and the output layer is the wavelength-intensity matrix; the matrix of the decoupled spectral component data is subjected to fiber lock-in amplification processing, and the peak value of the reflection spectrum is located to obtain decoupled fiber peak value data; and the decoupled fiber peak value data is subjected to spectral tuning to obtain decoupled spectral component data.

[0142] The dynamic time-frequency domain decoupling and feature extraction module is configured to perform dynamic time-frequency domain joint decoupling on the decoupled spectral component data, and to remove residual noise components to obtain a decoupled physical quantity feature matrix;

[0143] The model fusion and demodulation modeling module is configured to perform model parameter fusion based on the physical quantity feature matrix to obtain model parameter fusion data, to perform error compensation using the model parameter fusion data, and to construct a MEMS fiber multi-parameter demodulation model.

[0144] The performance evaluation and report generation module is configured to perform demodulation performance evaluation based on the MEMS fiber multi-parameter demodulation model to obtain fiber sensor demodulation performance evaluation data, and to generate a demodulation precision report using the fiber sensor demodulation performance evaluation data.

[0145] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, the scope of the application being defined by the appended claims and not by the above description, therefore all variations falling within the meaning and scope of the equivalent elements of the application file are intended to be included in the present application.

[0146] The above description is merely one specific implementation of the application, which enables a person skilled in the art to understand or implement the application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the application. Therefore, the application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A dynamic demodulation method for MEMS fiber multi-parameter sensors, characterized in that: The following steps are involved: Step S1: deploying a high-speed parallel acquisition sensor on the MEMS fiber multi-parameter sensor to collect spectral reflectance and spectral time domain data and construct an original multi-parameter spectral data set; Step S2: performing signal separation processing on the original multi-parameter spectral dataset based on a preset hybrid network to obtain a wavelength-intensity matrix, wherein the input layer based on the preset hybrid network is the original multi-parameter spectral dataset, and the output layer is the wavelength-intensity matrix; Performing optical fiber phase-locked amplification processing on the matrix of decoupled spectral component data and performing peak location of the reflection spectrum to obtain decoupled optical fiber peak data; performing spectral tuning on the decoupled optical fiber peak data to obtain decoupled spectral component data; Step S3: performing dynamic time-frequency domain joint decoupling on the decoupled spectral component data and removing the residual noise component to obtain a decoupled physical quantity characteristic matrix; Step S4: performing model parameter fusion based on the physical quantity characteristic matrix to obtain model parameter fusion data; Use model parameter fusion data to perform error compensation and build a MEMS fiber multi-parameter demodulation model; Step S5: performing demodulation performance evaluation based on the MEMS optical fiber multi-parameter demodulation model to obtain optical fiber sensor demodulation performance evaluation data; Generate a demodulation accuracy report using fiber optic sensor demodulation performance evaluation data.

2. The dynamic demodulation method for MEMS fiber multi-parameter sensor according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: deploying a multi-channel fiber Bragg grating array in the sensitive area of ​​the MEMS fiber sensor, with a wavelength coverage range of 1520-1570 nm, a temperature range of -200°C to 500°C, a strain range of 5000 με, a pressure range of 0-3.5 MPa, a sampling frequency of 8 MHz, and an accuracy of 0.1% FS, to collect the original spectral reflection signal; Step S12: using a photonic integrated ADC chip to perform time-interleaved sampling on the original spectral reflection signal, with a single-channel sampling rate of 250 kHz, 32 channels in parallel to achieve a total sampling rate of 8 MHz, and a quantization bit width of 18 bits, to generate multi-channel spectral time domain data; Step S13: calibrate the multi-channel spectral time domain data for the environmental noise baseline, and remove the electromagnetic interference pulses through the adaptive threshold method to generate denoised spectral time domain data; perform spectral reconstruction on the denoised spectral time domain data, and use non-uniform Fourier transform to convert the time domain signal into wavelength-intensity distribution to generate the original multi-parameter spectral data set.

3. The dynamic demodulation method for MEMS fiber multi-parameter sensor according to claim 1, characterized in that: Step S3 of performing dynamic time-frequency domain joint decoupling on the decoupled spectral component data includes: Adaptive noise covariance matrix is ​​used to track signal mutations of decoupled spectral component data to obtain spectral component tracking signals, where the update frequency of the adaptive noise covariance matrix is ​​1kHz; time domain processing is performed on the spectral component tracking signals to obtain time domain processing data of the optical fiber sensor; The decoupled spectral component data is sparsely reconstructed to obtain spectral component reconstructed data; the spectral component reconstructed data is frequency-domain processed to obtain spectral component frequency-domain processed data.

4. The dynamic demodulation method for MEMS fiber optic multi-parameter sensors according to claim 1, characterized in that: The residual noise component removal in step S3 includes: Get the noise dictionary library; The noise dictionary library is used to match and track the optical fiber sensor time domain processing data and the spectral component frequency domain processing data to obtain the optical fiber noise tracking data; Perform noise anomaly screening on the fiber noise tracking data and remove residual noise components to obtain fiber noise screening data; The signal regularity spectral density of the optical fiber noise screening data is calculated with the calculation standard being SNR greater than 35dB. The characteristic matrix is ​​then constructed to obtain the characteristic matrix of the decoupled physical quantity.

5. The dynamic demodulation method for MEMS fiber optic multi-parameter sensor according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: performing temperature-strain-pressure cross-sensitivity analysis on the physical quantity characteristic matrix and performing model parameter fusion to obtain model parameter fusion data; Step S42: Perform back propagation fine-tuning on the model parameter fusion data to obtain reverse parameter fine-tuning data; determine the decoupling accuracy based on the reverse parameter fine-tuning data. When the decoupling accuracy is less than 1.5%, convert the reverse parameter fine-tuning data into 8-bit fixed-point data and compress the parameter amount to less than 300,000 to obtain model lightweight parameter data; Step S43: constructing a MEMS fiber multi-parameter demodulation model using the model lightweight parameter data.

6. The dynamic demodulation method for MEMS fiber optic multi-parameter sensor according to claim 5, characterized in that: Step S41 includes the following steps: Step S411: performing normalization processing on the physical quantity characteristic matrix to obtain a physical standardized characteristic matrix; Step S412: performing temperature-strain-pressure covariance analysis and cross-sensitivity analysis using a physical standardized characteristic matrix to obtain a fiber cross-sensitivity coefficient matrix, wherein the off-diagonal elements of the fiber cross-sensitivity coefficient matrix represent temperature-strain data, temperature-pressure data, and strain-pressure data; Step S413: performing interference extraction and analysis on the temperature-strain data, the temperature-pressure data, and the strain-pressure data to obtain interference data of the optical fiber sensor; Step S414: using the interference data of the optical fiber sensor to perform model parameter fusion on the optical fiber cross sensitivity coefficient matrix according to the convergence condition of less than 0.5% and the number of iterations of 10 to obtain model parameter fusion data.

7. The dynamic demodulation method for MEMS fiber optic multi-parameter sensors according to claim 5, characterized in that: Step S43 includes the following steps: Step S431: converting the model lightweight parameter data into floating point format to obtain FPGA executable parameter data; Step S432: performing embedded logic unit mapping based on the FPGA executable parameter data to obtain optical fiber hardware model construction parameters; Step S433: Integrate the optical fiber hardware model construction parameters into the optical fiber sensor to construct a demodulation model, and obtain a MEMS optical fiber multi-parameter demodulation model.

8. The dynamic demodulation method for MEMS fiber optic multi-parameter sensors according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: constructing an evaluation calibration data set based on the MEMS fiber multi-parameter demodulation model; Step S52: performing demodulation performance evaluation using the evaluation and calibration data set to obtain optical fiber sensor demodulation performance evaluation data; Step S53: Using the optical fiber sensor demodulation performance evaluation data, the error distribution histogram is plotted according to the bin number of 20 and the 95% confidence interval is marked; using the optical fiber sensor demodulation performance evaluation data, the optical fiber response curve is generated according to the time resolution of 1 ms and superimposed with the original multi-parameter spectral data set; using the optical fiber sensor demodulation performance evaluation data, the demodulation value is calculated, and a temperature-strain-pressure consistency scatter plot is constructed to obtain a fiber consistency scatter plot; Step S54: Visualize and summarize the error distribution histogram, fiber response curve, and fiber consistency scatter plot, and output a demodulation accuracy report.

9. The dynamic demodulation method for MEMS fiber optic multi-parameter sensors according to claim 8, characterized in that: Step S52 includes the following steps: Step S521: Calculate the true value error of the demodulated value range of temperature-strain-pressure according to the evaluation calibration data set, and construct a matrix distribution to obtain a fiber error distribution matrix; Step S522: performing a step signal recording response time test on the evaluation calibration data set to obtain optical fiber response time test data; Step S523: extracting a demodulated signal based on the evaluation and calibration data set; performing power spectrum analysis on the demodulated signal to obtain demodulated signal evaluation data, wherein the frequency band of the power spectrum analysis is 1-0 kHz; Step S524: performing random weight data fusion on the optical fiber error distribution matrix, the optical fiber response time test data, and the demodulation signal evaluation data to obtain optical fiber sensor demodulation performance evaluation data.

10. A dynamic demodulation system for MEMS fiber optic multi-parameter sensors, characterized in that: The method for performing the dynamic demodulation method for MEMS fiber multi-parameter sensors according to claim 1, wherein the dynamic demodulation system based on the MEMS fiber multi-parameter sensor comprises: The raw data acquisition module is used to deploy high-speed parallel acquisition sensors for MEMS fiber optic multi-parameter sensors, collect spectral reflection and spectral time domain data, and construct a raw multi-parameter spectral data set; A signal decoupling and preprocessing module is used to perform signal separation processing on the original multi-parameter spectral data set based on a preset hybrid network to obtain a wavelength-intensity matrix, wherein the input layer based on the preset hybrid network is the original multi-parameter spectral data set, and the output layer is the wavelength-intensity matrix; the matrix of decoupled spectral component data is subjected to optical fiber phase-locked amplification processing and the peak location of the reflection spectrum is performed to obtain decoupled optical fiber peak data; the decoupled optical fiber peak data is subjected to spectral tuning to obtain decoupled spectral component data; Dynamic time-frequency domain decoupling and feature extraction module, used to perform dynamic time-frequency domain joint decoupling of decoupled spectral component data and remove residual noise components to obtain the decoupled physical quantity feature matrix; The model fusion and demodulation modeling module is used to fuse model parameters based on the physical quantity characteristic matrix to obtain model parameter fusion data; use the model parameter fusion data to perform error compensation and build a MEMS fiber multi-parameter demodulation model; The performance evaluation and report generation module is used to evaluate the demodulation performance based on the MEMS fiber multi-parameter demodulation model to obtain the fiber optic sensor demodulation performance evaluation data; and to generate a demodulation accuracy report using the fiber optic sensor demodulation performance evaluation data.

Images