A method and system for detecting humidity by misaligned fusion splicing mach-zehnder fiber

Abstract

The application discloses a misaligned fusion Mach-Zehnder fiber humidity detection method and system, relates to the field of optical fiber sensing and detection, and comprises the following steps: acquiring an original composite interference spectrum output by a misaligned fusion humidity sensor, inputting the original composite interference spectrum into a double-domain feature mapping network, and converting the original composite interference spectrum from a wavelength domain into an equivalent cavity length sequence in a spatial domain; extracting a parasitic high-frequency feature peak with an energy density located in a preset isolated interval in the sequence, inputting a mask model, replacing an energy density value corresponding to the feature peak with zero to output a denoising spatial domain sequence; performing reverse feature mapping conversion on the denoising sequence to return to the wavelength domain, and generating a pure interference spectrum; and extracting a wavelength offset of a target interference valley in the pure spectrum and mapping the wavelength offset into an environmental humidity value. The application solves the problem that a misaligned fusion structure introduces a parasitic high-frequency mode to cause interference spectrum distortion and then reduce humidity demodulation precision, and realizes accurate targeted denoising while retaining main interference characteristics.

Description

Technical Field

[0001] This invention relates to the field of fiber optic sensing and detection, and more specifically, to a method and system for humidity detection of misaligned fusion spliced ​​Mach-Zehnder fiber optic cables. Background Technology

[0002] Accurate monitoring of ambient humidity plays a crucial role in numerous fields, including industrial production, meteorological observation, agricultural storage, and the manufacturing of high-precision instruments. In recent years, fiber optic sensors have been widely used in humidity detection due to their inherent advantages such as resistance to electromagnetic interference, small size, corrosion resistance, and ease of distributed measurement. Among them, the Mach-Zehnder interferometer fiber optic sensor based on a staggered fusion splice structure is a highly representative structure. Its basic working principle involves artificially creating a small lateral physical misalignment at the connection point of two fiber segments, allowing some of the light originally propagating in the fiber core to couple into the fiber cladding. This cladding light, during its forward propagation, interacts with the evanescent field of the humidity-sensitive material coated on the cladding, and then re-interferes with the core light at the next staggered fusion splice or coupling point. When the ambient humidity changes, the refractive index or volume of the humidity-sensitive coating changes accordingly, causing a regular shift in the interference valley wavelength of the interference spectrum. By demodulating this shift, the ambient humidity can be calculated.

[0003] In practical engineering applications and long-term monitoring, this sensing mechanism faces extremely challenging multi-faceted interference problems. First, the misaligned welding process is physically rough and asymmetrical. This microscopic geometric abruptness, while exciting the dominant interference mode, inevitably excites higher-order parasitic cladding modes or end-face reflection modes. These parasitic modes manifest as high-frequency spikes and distortions superimposed on the main interference envelope in the spectrum. Traditional signal processing methods often rely on simple smoothing filters or frequency domain low-pass filters. However, while filtering out these high-frequency noises, they easily weaken the characteristic depth of the true interference valleys and even cause spurious wavelength shifts, directly limiting the absolute accuracy of humidity demodulation.

[0004] In actual operation, sensors are often exposed to high humidity condensation or complex atmospheric environments. When the environment undergoes a receding phase transition or becomes supersaturated with humidity, tiny droplets easily condense on the surface of the optical fiber. These transient droplets form random, localized strong scattering centers, disrupting the transmission conditions of specific optical modes and causing the originally clear interference valleys to suddenly become shallower or distorted. Chemical impurities suspended in the air over a long period can also slowly penetrate into the humidity-sensitive coating, causing irreversible drift in the coating's base refractive index. This long-term, slowly varying drift caused by impurity penetration is deeply coupled with the actual humidity change signal. Conventional single-wavelength tracking techniques cannot discern the true physical cause of the signal drift, leading to severe baseline drift and measurement distortion in the sensor after long-term service.

[0005] To address the complex interference spectral degradation issues caused by structural physical defects, transient phase transition scattering, and long-term infiltration of heterogeneous substances, the industry urgently needs a novel method for spectral analysis and feature decoupling to overcome the current limitations in the accuracy and lifespan of humidity detection. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and system for humidity detection of misaligned fusion spliced ​​Mach-Zehnder optical fibers, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for humidity detection of misaligned fusion spliced ​​Mach-Zehnder optical fiber includes: Obtain the original composite interference spectrum output by the misaligned fusion spliced ​​Mach-Zehnder fiber optic humidity sensor; The original composite interference spectrum is input into a pre-trained dual-domain feature mapping network, which converts the original composite interference spectrum from the wavelength domain into a spatial domain equivalent cavity length sequence. Parasitic high-frequency feature peaks with energy density located within a preset isolated interval are extracted from the spatial domain equivalent cavity length sequence. The spatial domain equivalent cavity length sequence and the parasitic high-frequency feature peaks are input into the masking model. The masking model replaces the energy density values ​​corresponding to the parasitic high-frequency feature peaks with zero and outputs a denoised spatial domain sequence. Perform inverse feature mapping on the denoised spatial domain sequence to transform the denoised spatial domain sequence back from the spatial domain to the wavelength domain, generating a pure Mach-Zehnder interferometer spectrum; The wavelength shift of the target interference valley is extracted from the pure Mach-Zehnder interferometer spectrum, and the wavelength shift is mapped to the ambient humidity value.

[0008] Furthermore, before inputting the original composite interference spectrum into the pre-trained dual-domain feature mapping network, the steady-state interference spectrum of the misaligned fusion spliced ​​Mach-Zehnder fiber humidity sensor under humidity-free conditions is obtained as sample input data. At the same time, the ideal interference spectrum determined by the interferometer structural parameters is obtained as sample label data. The initial dual-domain feature mapping network is iteratively optimized using the sample input data and sample label data until the mean square error between the output data of the dual-domain feature mapping network and the sample label data converges to within the error convergence condition, thus completing the training of the dual-domain feature mapping network.

[0009] Furthermore, the step of extracting parasitic high-frequency characteristic peaks with energy density located within a preset isolated interval from the spatial domain equivalent cavity length sequence includes: The spatial domain equivalent cavity length sequence is input into the peak isolation model. The peak isolation model divides the spatial domain equivalent cavity length sequence into a low-frequency energy accumulation region and a high-frequency energy accumulation region, and outputs the isolated peaks that are in the high-frequency energy accumulation region and whose cavity length values ​​are within the preset isolation interval as the parasitic high-frequency characteristic peaks.

[0010] Furthermore, after generating a pure Mach-Zehnder interferometer spectrum, the peak energy density and central cavity length of the parasitic high-frequency characteristic peaks are extracted. The peak energy density and central cavity length are input into the aging assessment model. The aging assessment model maps the peak energy density and central cavity length to the coating interface peeling aging index, outputs the coating interface peeling aging index, and triggers an early warning signal.

[0011] Furthermore, in the step of inputting the spatial domain equivalent cavity length sequence and the parasitic high-frequency feature peak into the masking model, the masking model generates a binarized matrix with the same dimension as the spatial domain equivalent cavity length sequence. The matrix element values ​​of the binarized matrix at the cavity length position corresponding to the parasitic high-frequency feature peak are zero, and the matrix element values ​​of the binarized matrix at the cavity length positions other than the cavity length position corresponding to the parasitic high-frequency feature peak are one. The masking model performs a matrix dot product operation between the binarized matrix and the spatial domain equivalent cavity length sequence to generate the noise-reduced spatial domain sequence.

[0012] Furthermore, after generating a pure Mach-Zehnder interferometer spectrum, the interference valley sequence corresponding to multiple adjacent cladding modes in the pure Mach-Zehnder interferometer spectrum is obtained. Within a predetermined time period, the depth contrast value and center wavelength value of each interference valley in the interference valley sequence are continuously extracted to generate a multi-valley contrast time series and a multi-valley center wavelength time series. The multi-valley contrast time series and the multi-valley center wavelength time series are input into the transient droplet scattering identification model. The transient droplet scattering identification model processes the contrast attenuation change of each interference valley through an attention mechanism and outputs the failure probability weight corresponding to each interference valley. The dynamic routing matrix is ​​calculated by taking the failure probability weights of all interference valleys and the time sequence input wavelength of the multi-valley center wavelength. The dynamic routing matrix is ​​used to determine the interference valleys in the time sequence input wavelength of the multi-valley center wavelengths where the failure probability weights have not triggered the routing truncation condition as the target interference valleys.

[0013] Furthermore, before inputting the multi-valley contrast time series and multi-valley center wavelength time series into the transient droplet scattering recognition model, the multi-valley contrast time series and multi-valley center wavelength time series under the receding phase transition state are collected as positive sample data, and the multi-valley contrast time series and multi-valley center wavelength time series under the stable environment state are collected as negative sample data. The initial transient droplet scattering recognition model is trained using the positive sample data and the negative sample data. During the training process, the attention weight node parameters inside the transient droplet scattering recognition model are adjusted. When the failure probability weight output by the transient droplet scattering recognition model for the positive sample data matches the manually labeled data, the training of the transient droplet scattering recognition model ends.

[0014] Furthermore, in the step of mapping the wavelength offset to an ambient humidity value: The wavelength drift trajectory sequence of at least three independent interference valleys of different orders in the pure Mach-Zehnder interferometer spectrum is obtained in multiple time periods. The wavelength drift trajectory sequence is input into the heterogeneous material diffusion gradient analytical model. The heterogeneous material diffusion gradient analytical model extracts the phase difference feature and asynchronous drift rate feature of the independent interference valleys of different orders in the wavelength drift trajectory sequence through a graph neural network. The phase difference feature and asynchronous drift rate feature are mapped to the impurity penetration equivalent refractive index deviation value. The wavelength offset of the target interference valley is subtracted from the equivalent refractive index deviation of the impurity penetration to generate the calibrated wavelength offset, which is then mapped to the ambient humidity value.

[0015] Furthermore, in the heterogeneous material diffusion gradient analytical model, the graph neural network constructs at least three independent interference valleys as three independent nodes of the graph structure. The graph neural network initializes the edge weights between the three independent nodes to one. During the feature extraction process, the wavelength evolution time lag between the three independent nodes is used as the update compensation term for the edge weights. The refractive index gradual mapping value is calculated using the update compensation term, and the refractive index gradual mapping value is used as the impurity penetration equivalent refractive index deviation value.

[0016] This invention also discloses a humidity detection system for misaligned fusion spliced ​​Mach-Zehnder fiber optic cables to implement the above method, the humidity detection system comprising: The spectrum acquisition module is used to acquire the original composite interference spectrum output by the misaligned fusion spliced ​​Mach-Zehnder fiber optic humidity sensor. The dual-domain mapping module is used to input the original composite interference spectrum into a pre-trained dual-domain feature mapping network, which converts the original composite interference spectrum from the wavelength domain into a spatial domain equivalent cavity length sequence. The parasitic masking module is used to extract parasitic high-frequency feature peaks with energy density located in a preset isolated interval from the spatial domain equivalent cavity length sequence. The spatial domain equivalent cavity length sequence and the parasitic high-frequency feature peaks are input into the masking model. The masking model replaces the energy density value corresponding to the parasitic high-frequency feature peak with zero and outputs a denoised spatial domain sequence. The spectral reconstruction module is used to perform inverse feature mapping on the denoised spatial domain sequence, converting the denoised spatial domain sequence from the spatial domain back to the wavelength domain, and generating a pure Mach-Zehnder interferometer spectrum. The humidity demodulation module is used to extract the wavelength shift of the target interference valley in the pure Mach-Zehnder interferometer spectrum and map the wavelength shift to the ambient humidity value.

[0017] The advantages of this invention over the prior art are: This invention effectively solves the problem of high-frequency spectral distortion caused by inherent physical defects in misaligned fusion structures, significantly improving the wavelength demodulation accuracy of the target interference valley. Instead of employing traditional mathematical filtering methods that easily lead to signal distortion, this invention innovatively introduces a pre-trained dual-domain feature mapping network. This network can losslessly convert the original composite interference spectrum affected by parasitic modes into a spatial domain equivalent cavity length sequence. Utilizing the physical law that parasitic high-frequency modes appear as isolated energy accumulation regions in the spatial domain, this invention uses a masking model to zero out the energy density at specific cavity length locations, followed by inverse feature mapping. This targeted noise reduction method based on precise physical localization in the spatial domain removes parasitic high-frequency spikes while preserving the spectral morphology of the main interference valley, fundamentally eliminating the systematic errors caused by peak-finding algorithms.

[0018] This invention also effectively solves the measurement jump problem caused by the condensation of tiny droplets in high humidity environments or during the receding phase transition, endowing the system with extremely strong resistance to transient scattering interference. This invention is not limited to tracking a single interference valley, but extracts interference valley sequences corresponding to multiple adjacent cladding modes and introduces a transient microdroplet scattering identification model. This model utilizes an attention mechanism to deeply mine the dynamic law of the non-natural attenuation of interference valley contrast, accurately quantifying and outputting the failure probability weight of each interference valley affected by droplet scattering. Combined with wavelength calculation of a dynamic routing matrix, the system can truncate the interfered failure channels in real time and dynamically route the demodulation calculation to interference valleys in a healthy state. This mechanism of dynamic fault tolerance using multimodal spatial distribution differences avoids the drastic fluctuations in humidity data under condensation conditions.

[0019] This invention also effectively solves the baseline drift problem caused by long-term penetration of heterogeneous chemical impurities in complex environments. Utilizing the physical principle that water molecules and impurity molecules diffuse at different rates within the coating, this invention tracks at least three independent interference valleys at different orders using an analytical model of the diffusion gradient of heterogeneous substances. The graph neural network within the model constructs these independent interference valleys as network nodes, using the time lag of wavelength evolution between different modes as an update compensation term for the edge weights. This characteristic approach allows the system to accurately extract the asynchronous drift rate and phase difference caused by impurity penetration from the seemingly chaotic wavelength drift trajectory, and then calculate the equivalent refractive index deviation of the impurity penetration for reverse compensation. This not only purifies the true humidity signal but also significantly extends the calibration-free service life of the sensor in harsh atmospheric environments. Attached Figure Description

[0020] Figure 1 This is the core method flow and signal state evolution diagram of the present invention, reflecting the overall logic of the spectrum being converted from the wavelength domain to the spatial domain, then undergoing masking processing, and finally returning to the pure spectrum to generate humidity data.

[0021] Figure 2 This is a schematic diagram of the actual structure of the staggered fusion splice Mach-Zehnder fiber of the present invention, showing the relationship between the input fiber, the staggered fusion splice, the moisture-absorbing coating sensing area, and the cladding / core interference optical path.

[0022] Figure 3 This invention presents a peak isolation and spatial domain equivalent cavity length sequence analysis diagram. It uses a mathematical coordinate system to draw the division of low-frequency and high-frequency energy accumulation regions, intuitively demonstrating the extraction and location of parasitic high-frequency characteristic peaks.

[0023] Figure 4 This is a screenshot of the binary matrix multiplication operation of the masking model of the present invention. It uses waveform multiplication to illustrate how to use a binary matrix with zeros set at the corresponding cavity length position to flatten parasitic peaks and output a noise reduction sequence.

[0024] Figure 5 This is a diagram of the coating aging assessment mechanism of the present invention, which combines the microstructure of the peeling phenomenon with the aging assessment model process, demonstrating the mechanism of triggering early warning based on characteristic peak parameters.

[0025] Figure 6 This invention presents a logic diagram for multi-valley contrast and transient droplet identification, showing the temporal evolution of three-order interference valleys and how to cut off disturbed interference valleys and filter out target interference valleys through attention mechanisms and dynamic routing matrices.

[0026] Figure 7This invention relates to a graph neural network structure and a graph for extracting features of heterogeneous materials. It visually demonstrates how three independent interference valleys are constructed as GNN nodes, and how time lag is used as an update mechanism for edge weights to ultimately extract the features of material diffusion.

[0027] Figure 8 This is the final humidity calibration and demodulation relationship diagram of the present invention. Through the relative relationship between two nonlinear curves in the coordinate system, it shows the mathematical calibration process of subtracting the refractive index deviation from the disturbed original drift amount and finally mapping it to the actual calibrated humidity. Detailed Implementation

[0028] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0029] This embodiment provides a method for humidity detection using misaligned fusion spliced ​​Mach-Zehnder fiber optic cables. This method is applicable to a detection platform consisting of a broadband light source, a misaligned fusion spliced ​​Mach-Zehnder fiber optic humidity sensor, a spectral acquisition unit, and a data processing unit. The broadband light source inputs broadband light into the fiber optic cable, and the misaligned fusion spliced ​​Mach-Zehnder fiber optic humidity sensor outputs an interference spectrum. The spectral acquisition unit can be a spectrometer, a fiber optic demodulator, or a photoelectric detection component with a beam splitter and analog-to-digital conversion interface. The data processing unit can be implemented using a host computer, an embedded processor, a field-programmable gate array (FPGA), or an edge computing module.

[0030] A staggered fusion splice Mach-Zehnder fiber optic humidity sensor may include an input fiber, an output fiber, two staggered fusion splices, and a moisture-absorbing coated sensing area located between the two staggered fusion splices. Figure 2 As shown, after the input light enters the first misaligned fusion point, due to the lateral offset between the fiber cores, part of the light continues to propagate along the fiber core, while the other part is coupled into the cladding and forms a cladding mode. The core light and the cladding light have different effective refractive indices and optical path lengths within the sensing area, and the cladding light also interacts with the evanescent field of the hygroscopic coating. When the external humidity changes, the refractive index, water content, or expansion state of the hygroscopic coating changes, and the phase of the cladding mode changes accordingly. It then recouples with the core light near the second misaligned fusion point, generating Mach-Zehnder interference. By tracking the shift in the center wavelength of the interference valley, the ambient humidity can be obtained.

[0031] In actual testing, the original composite interference spectrum output by the misaligned fusion spliced ​​Mach-Zehnder fiber humidity sensor is first acquired. The original composite interference spectrum generally includes the main Mach-Zehnder interference envelope, parasitic cladding modes, higher-order reflection modes, and acquisition noise. To avoid weakening the true interference valleys using traditional smoothing filters, this invention does not directly and coarsely filter out high-frequency fluctuations in the wavelength domain. Instead, the original composite interference spectrum is input into a pre-trained dual-domain feature mapping network. The dual-domain feature mapping network converts the original composite interference spectrum from the wavelength domain into a spatial domain equivalent cavity length sequence, such as... Figure 1As shown.

[0032] The spatial domain equivalent cavity length sequence here refers to converting interference fluctuations at different frequencies in the spectrum into energy density distributions at different equivalent optical path differences or equivalent cavity length positions. In practical processing, the original composite interference spectrum can first be normalized in intensity and calibrated on the wavelength axis, and then the wavelength sampling data can be resampled to equally spaced wavenumber coordinates to maintain a stable correspondence between the interference period and the equivalent optical path difference. The energy corresponding to the principal Mach-Zehnder interference is usually concentrated in a low-frequency energy accumulation region that matches the effective length of the sensing area, while parasitic high-frequency interference caused by misaligned fusion ends, local peeling interfaces, and higher-order cladding modes often presents as discrete, narrow, and relatively isolated energy peaks in the spatial domain. Therefore, this invention transforms the glitch interference, which is difficult to accurately separate in the wavelength domain, into the problem of processing isolated peaks whose positions can be identified in the spatial domain.

[0033] A dual-domain feature mapping network can be constructed using a 1D convolutional encoder and a 1D deconvolutional decoder. The 1D convolutional encoder extracts local oscillation features and envelope features from the original composite interference spectrum, while the deconvolutional decoder outputs the energy density sequence corresponding to the sampling points of the equivalent cavity length. To preserve the physical meaning of the spatial domain, a Fast Fourier Transform (FFT) layer or a Discrete Cosine Transform (DCT) layer can be set as a fixed feature layer in the network to transform the resampled intensity sequence to the feature space related to the equivalent optical path difference. Then, a trainable convolutional layer performs error compensation on the fixed transformation result. The purpose of this setup is to utilize both the physical interpretability of traditional spectral transformations and the neural network to correct deviations caused by non-ideal misaligned weld structures, non-uniform sampling, and coating dispersion.

[0034] After generating the equivalent cavity length sequence in the spatial domain, parasitic high-frequency characteristic peaks with energy density located within a preset isolated interval are extracted from this sequence. The preset isolated interval can be obtained based on the sensor's factory calibration. Specifically, under constant humidity, no condensation, and stable coating conditions, the sensor is repeatedly sampled, and the cavity length range where the main Mach-Zehnder energy peak is located is statistically analyzed. This range is taken as the low-frequency energy accumulation region. Then, the positions where parasitic spikes repeatedly appear in the spatial domain are statistically analyzed, and the cavity length range that is far from the low-frequency energy accumulation region and has a narrow peak width is taken as the preset isolated interval. The lower boundary of the preset isolated interval can be set to 1.1 to 3 times the upper boundary of the low-frequency energy accumulation region, and the upper boundary can be set to 60% to 95% of the maximum resolvable equivalent cavity length of the system. Since the misalignment, coating thickness, and effective sensing length of different sensors vary, the preset isolated interval is preferably further narrowed according to the equivalent cavity length coordinates during the product calibration stage.

[0035] The extraction of parasitic high-frequency characteristic peaks can be accomplished using a peak isolation model. After receiving the spatial domain equivalent cavity length sequence, the peak isolation model divides the sequence into low-frequency energy accumulation regions and high-frequency energy accumulation regions based on the energy density distribution, such as... Figure 3As shown. Low-frequency energy accumulation regions typically contain main interference information and should not be directly set to zero. In high-frequency energy accumulation regions, only isolated peaks that meet the following conditions are output as parasitic high-frequency characteristic peaks: their cavity length falls within a preset isolated interval, their peak energy density is higher than the local background energy, and their peak width is smaller than the peak width of the main interference peak. The local background energy can be estimated using the median energy of adjacent windows. The ratio of peak energy density to local background energy can be set to 1.5 to 10, and the peak width can be set to 0.05 to 0.4 times the peak width of the main interference peak. These constraints can prevent the main Mach-Zehnder interference peak or the true humidity response from being misjudged as parasitic interference.

[0036] Subsequently, the spatial domain equivalent cavity length sequence and the extracted parasitic high-frequency feature peaks are input into the masking model. The masking model does not perform low-pass filtering on the entire signal, but rather performs point-to-point suppression only at the cavity length locations where the parasitic high-frequency feature peaks are located in the spatial domain. Specifically, the masking model generates a binary matrix with the same dimensions as the spatial domain equivalent cavity length sequence, such as... Figure 4 As shown, the binarization matrix is ​​set to 0 at the cavity length corresponding to the parasitic high-frequency feature peak and 1 at other cavity length positions. To prevent residual edge energy on both sides of the parasitic peak, the positions where 0 is set can cover the cavity length of the center of the parasitic peak and several adjacent sampling points on the left and right, and the coverage width can be 1 to 3 times the half-width at half-maximum of the parasitic peak. The masking model then performs a matrix multiplication operation between the binarization matrix and the equivalent cavity length sequence in the spatial domain. The energy density of the parasitic high-frequency feature peak is replaced with 0, while the energy of the other spatial domains remains unchanged, and the final output is a denoised spatial domain sequence.

[0037] Inverse feature mapping is performed on the denoised spatial domain sequence to transform it back into the wavelength domain, generating a pure Mach-Zehnder interferometric spectrum. Inverse feature mapping can be implemented using the inverse decoding branch in a dual-domain feature mapping network, or it can be implemented using an inverse Fourier transform, inverse discrete cosine transform, or a trainable deconvolutional network that matches the forward spatial domain transformation. When the forward mapping includes a neural network compensation layer, the inverse mapping can also be configured with a corresponding inverse compensation layer to reduce spectral reconstruction errors. The reconstructed pure Mach-Zehnder interferometric spectrum retains the depth, width, and center wavelength of the main interference valleys while removing fine spikes caused by parasitic high-frequency modes. Therefore, subsequent valley-finding results are less susceptible to being influenced by local distortions.

[0038] The wavelength shift of the target interference valley is extracted from the pure Mach-Zehnder interferometry spectrum and mapped to the ambient humidity value. The target interference valley can be the main interference valley selected during the calibration phase or a healthy interference valley selected by the subsequent dynamic routing matrix. The extraction of the center wavelength of the interference valley can be achieved using quadratic curve fitting, Gaussian fitting, centroid method, or local minimum interpolation. The wavelength shift is the change in the center wavelength of the current target interference valley relative to the center wavelength under the reference humidity. Humidity mapping can be completed using calibration curves, lookup table interpolation, or regression models. The calibration curve can be established from multiple known relative humidity points, for example, segmented calibration within the range of 10%RH to 95%RH. The humidity resolution can be set to 0.1%RH to 1%RH based on the spectrometer sampling interval and sensor sensitivity.

[0039] In a further embodiment, network training is required before the dual-domain feature mapping network is used for actual detection. During training, the steady-state interference spectrum of the misaligned fusion spliced ​​Mach-Zehnder fiber optic humidity sensor under humidity-undisturbed conditions is acquired as sample input data. A humidity-undisturbed state can be understood as a state with constant ambient humidity, no condensation, no rapid airflow disturbance, and a stable coating condition, with humidity fluctuations controlled within 0.1%RH to 2%RH. Simultaneously, an ideal interference spectrum is determined based on the interferometer's structural parameters as sample label data. Structural parameters may include the sensing zone length between the two misaligned fusion splices, the effective refractive index of the fiber core, the effective refractive index of the cladding, the misalignment amount, the coating thickness, and the coating refractive index. The ideal interference spectrum can be obtained through simulation using a Mach-Zehnder interferometry propagation model or generated from a reference spectrum after high-precision calibration and multiple averaging.

[0040] The training process involves feeding the sample input data into an initial dual-domain feature mapping network. The network first generates a spatial domain equivalent cavity length sequence, then generates a reconstructed interference spectrum through an inverse reconstruction branch. The reconstructed interference spectrum is then compared with the sample label data based on mean square error. This process ensures that both the network output and the sample labels are in wavelength domain spectral form, avoiding dimensional inconsistencies caused by directly comparing the spatial domain sequence with the ideal interference spectrum. During training, auxiliary constraints can be applied to the intermediate spatial domain equivalent cavity length sequence to concentrate the main interference energy in the low-frequency energy accumulation region and suppress abnormal peaks within a preset isolated interval. Iterative optimization continues until the mean square error between the dual-domain feature mapping network output data and the sample label data converges to the error convergence condition. The error convergence condition can be set as a mean square error change rate of less than 0.1% to 1% over 10 to 50 consecutive training rounds, or a mean square error lower than a preset calibration error upper limit. To prevent the network from learning parasitic high-frequency modes as effective features, data with different misalignment amounts, coating thicknesses, and sampling noise conditions can be added to the training set, making the network more inclined to retain stable main Mach-Zehnder interference features.

[0041] In a further embodiment, after generating the pure Mach-Zehnder interferometry spectrum, the parasitic high-frequency characteristic peaks identified in the aforementioned steps can also be used for coating interface peeling aging assessment, such as... Figure 5 As shown. Although parasitic high-frequency characteristic peaks are detrimental to humidity demodulation, their peak energy density and central cavity length often reflect the additional reflection paths after minor peeling at the coating interface, weld end, or cladding boundary. When local peeling, cracking, or interface voids occur in the hygroscopic coating, the new optical interface will enhance local reflection or scattering, thereby causing the energy of some isolated peaks in the spatial domain to gradually increase or their positions to slowly shift.

[0042] In practice, the data processing unit extracts the peak energy density and central cavity length of the parasitic high-frequency characteristic peaks and inputs them into the aging assessment model. The aging assessment model can be implemented using a linear regression model, support vector regression model, random forest regression model, or a small multilayer perceptron. The model outputs a coating interface peeling aging index, which can be normalized to between 0 and 1. A value close to 0 indicates a stable coating interface state, while a value close to 1 indicates a high risk of coating peeling. A warning signal can be triggered when the aging index is greater than 0.6 to 0.9, or when the aging index continues to rise for 3 to 20 consecutive detection cycles. This is because a single high-energy parasitic peak may be caused by transient perturbations, while a continuously increasing parasitic peak better reflects irreversible coating aging.

[0043] In a further embodiment, after generating a pure Mach-Zehnder interferometry spectrum, multi-valley dynamic fault-tolerant demodulation can also be performed, such as... Figure 6 As shown. Specifically, the sequence of interference valleys corresponding to multiple adjacent cladding modes in the pure Mach-Zehnder interferometry spectrum is obtained. The number of interference valleys can be 3 to 10, preferably 3 to 6, to balance fault tolerance and computational load. The depth contrast value and center wavelength value of each interference valley are continuously extracted within a predetermined time period to generate a multi-valley contrast time series and a multi-valley center wavelength time series. The predetermined time period can be 5 s to 30 min; if used for monitoring condensation and receding phase transitions, it can be 5 s to 120 s; if used for long-term monitoring in ordinary environments, it can be 1 min to 30 min.

[0044] The depth contrast of an interference valley can be calculated from the difference or ratio between the local peak intensity on both sides of the valley and the intensity at the valley bottom. The center wavelength can be obtained through local fitting. If an interference valley is affected by scattering from tiny droplets, it will usually exhibit unnatural contrast attenuation; that is, changes in humidity are insufficient to explain the sudden shallowing of the valley depth, while the center wavelength may fluctuate abnormally. Based on this characteristic, the multi-valley contrast time series and the multi-valley center wavelength time series are input into the transient droplet scattering identification model.

[0045] The transient droplet scattering identification model can be implemented using temporal convolutional networks, long short-term memory networks, or Transformer encoders, with an attention mechanism incorporated within the model. This attention mechanism assigns weights to different interference valleys and the contrast attenuation changes at different times, thereby identifying which interference valleys are more likely to be disrupted by local droplet scattering. Model inputs can include the contrast change, center wavelength change, drift consistency between adjacent interference valleys, and contrast recovery rate for each interference valley within a predetermined time period. The model outputs a failure probability weight for each interference valley, which can be normalized to between 0 and 1. A higher value indicates that the interference valley is less likely to be suitable for participating in the current humidity demodulation.

[0046] Subsequently, a dynamic routing matrix is ​​calculated using the failure probability weights of all interferometric valleys and the time-series input wavelengths of the multi-valley center wavelengths. The dynamic routing matrix can be understood as a selection matrix that chooses wavelength calculation channels based on health status, or it can be implemented as a gating matrix consistent with the number of interferometric valleys. If the failure probability weight of an interferometric valley triggers a routing truncation condition, the wavelength calculation channel corresponding to that valley is closed; if the routing truncation condition is not triggered, the interferometric valley is retained as a candidate channel. The routing truncation condition can be set to a failure probability weight greater than 0.5 to 0.9, or a failure probability weight greater than 0.5 to 0.9 within 2 to 10 consecutive sampling points. The dynamic routing matrix selects target interferometric valleys from those that have not been truncated, prioritizing those with the highest contrast, smoothest historical drift, and highest calibration sensitivity. In this way, even if a local droplet causes a valley to fail temporarily, the system can still switch to other healthy interferometric valleys, avoiding sudden jumps in output humidity.

[0047] In a further embodiment, the transient droplet scattering recognition model needs to be trained before use. The training samples include positive sample data and negative sample data. The positive sample data is collected from the multi-valley contrast time series and multi-valley center wavelength time series under the dehydration phase transition state. The dehydration phase transition state can be obtained by subjecting the sensor surface to high humidity, condensation, droplet evaporation, or wet film rupture. At this time, some interference valleys will show sudden contrast decay or inconsistent drift. The negative sample data is collected from the multi-valley contrast time series and multi-valley center wavelength time series under the stable environmental state. Under the stable environmental state, the humidity changes slowly and there is no obvious droplet scattering, and the center wavelength drift of multiple interference valleys has good consistency.

[0048] During training, each interference valley is manually labeled to indicate whether it is affected by transient droplet scattering within each time segment, and the labeling results are converted into target failure probability values. The initial transient droplet scattering identification model receives positive and negative sample data and outputs the failure probability weights for each interference valley. During backpropagation, the attention weight node parameters within the model are adjusted to gradually focus attention on key segments such as sudden contrast decay, delayed valley depth recovery, and inconsistent drift between adjacent valleys. Training ends when the failure probability weights output by the model for positive sample data match the manually labeled values. Matching conditions can be set to a classification accuracy greater than 90% to 99%, or an average failure probability greater than 0.7 for positive sample failed valleys and an average failure probability less than 0.3 for healthy valleys.

[0049] In a further embodiment, to address the baseline drift problem caused by long-term infiltration of heterogeneous chemical impurities in complex environments, the present invention also compensates for the humidity mapping process. Specifically, wavelength drift trajectory sequences of at least three independent interference valleys of different orders in the pure Mach-Zehnder interferometry spectrum are obtained over multiple time periods. These multiple time periods can range from 10 to 10,000 sampling periods; short-term analysis can use 10 to 300 sampling periods, while long-term aging analysis can use 300 to 10,000 sampling periods. The independent interference valleys of different orders originate from different cladding modes or different modal coupling orders, and their response rates to water molecule diffusion and impurity molecule infiltration are not entirely the same.

[0050] The wavelength drift trajectory sequence is input into the analytical model of diffusion gradient of heterogeneous materials. This model extracts the phase difference and asynchronous drift rate features of independent interference valleys of different orders in the wavelength drift trajectory sequence using a graph neural network. The phase difference feature can be understood as the difference in the timing of the responses of different interference valleys to the same humidity change, while the asynchronous drift rate feature can be understood as the degree of inconsistency in the drift velocity of each interference valley during long-term changes. The humidity response caused by water molecules usually exhibits a relatively fast and consistent reversible change, while the penetration of impurity molecules into the coating is slower and is subject to hysteresis and irreversible shifts due to differences in coating depth, local porosity, and evanescent field distribution of different modes. Therefore, the model estimates the equivalent refractive index deviation of impurity penetration through these temporal differences.

[0051] To avoid dimensional inconsistencies, the impurity penetration equivalent refractive index deviation is converted into an equivalent wavelength compensation value with the same dimensions as the target interference valley wavelength shift before being used for humidity demodulation. In the data processing flow, this equivalent wavelength compensation value can also be directly output as the impurity penetration equivalent refractive index deviation value. Subsequently, the impurity penetration equivalent refractive index deviation value is subtracted from the target interference valley wavelength shift to generate the calibrated wavelength shift, which is then mapped to the ambient humidity value. The compensation value range can be set to 0% to 30% of the full-scale wavelength drift of the target interference valley based on long-term calibration. When long-term contamination is relatively light, the compensation value is close to 0; when there is significant chemical penetration or baseline drift in the coating, the compensation value increases. Through this compensation process, the system output is closer to the wavelength change caused by actual humidity, rather than mistaking the slow drift caused by coating contamination for humidity changes. Figure 8 As shown.

[0052] In a further embodiment, the graph neural network structure in the analytical model of diffusion gradient of heterogeneous substances is as follows: Figure 7 As shown, the graph neural network constructs at least three independent interference valleys as at least three independent nodes in the graph structure. Taking three independent interference valleys as an example, three independent nodes are formed in the graph structure; when four or more independent interference valleys are selected, four or more nodes are formed accordingly. The input features of each node can include the center wavelength sequence, drift velocity sequence, drift acceleration sequence, and contrast stability of the interference valley over multiple time periods. Edges are set between the independent nodes, representing the correlation between different modal responses. The initial weight of the edges is 1, indicating that different interference valleys have equal correlation strength by default in the early stage of training.

[0053] During feature extraction, the graph neural network calculates the time lag of wavelength evolution between different nodes and uses this time lag as an update compensation term for the edge weights. The time lag can be obtained through the peak position of cross-correlation, dynamic time warping distance, or temporal attention alignment results. If the wavelength trajectories of two interference valleys change synchronously, it indicates that they are mainly driven by real humidity, and the edge weights remain high. If there is a significant time lag or drift velocity difference between the two interference valleys, it indicates that there is an asynchronous slow-change component caused by the infiltration of heterogeneous substances, and the edge weights are updated and compensated according to the lag. After multiple layers of message passing, the graph neural network aggregates the phase difference features, asynchronous drift rate features, and lag compensation features of each node to calculate the refractive index gradual change mapping value. This refractive index gradual change mapping value is then converted into a compensation value with the same dimensions as the wavelength offset of the target interference valley and used as the equivalent refractive index deviation value of impurity infiltration in subsequent calibration.

[0054] The present invention also provides a misaligned fusion spliced ​​Mach-Zehnder fiber optic humidity detection system for implementing the above method. The system includes a spectrum acquisition module, a dual-domain mapping module, a parasitic masking module, a spectrum reconstruction module, and a humidity demodulation module.

[0055] The spectral acquisition module is used to acquire the raw composite interference spectrum output by the misaligned fusion-bonded Mach-Zehnder fiber optic humidity sensor. This module may include a broadband light source drive unit, a spectrometer interface, a data buffer unit, and a spectral preprocessing unit. The spectral preprocessing unit can perform dark current subtraction, intensity normalization, outlier removal, wavelength axis alignment, and wavenumber axis resampling, but does not perform strong smoothing that would alter the center position of the interference valleys.

[0056] The dual-domain mapping module is used to input the raw composite interference spectrum into a pre-trained dual-domain feature mapping network and output a spatial domain equivalent cavity length sequence. This module can be deployed in an embedded graphics processor, digital signal processor, or host computer software. The dual-domain mapping module saves the trained network parameters, input normalization parameters, and cavity length calibration parameters, enabling different batches of sensors to generate corresponding spatial domain sequences based on their respective structural parameters.

[0057] The parasitic masking module extracts parasitic high-frequency characteristic peaks with energy densities located within a preset isolated interval from the spatial domain equivalent cavity length sequence, and inputs the spatial domain equivalent cavity length sequence and the parasitic high-frequency characteristic peaks into the masking model. The masking model replaces the energy density values ​​corresponding to the parasitic high-frequency characteristic peaks with 0, outputting a denoised spatial domain sequence. The parasitic masking module can also record the peak energy density, central cavity length, and occurrence frequency of the parasitic peaks, providing input for coating aging assessment.

[0058] The spectral reconstruction module performs inverse feature mapping on the denoised spatial domain sequence, transforming it back into the wavelength domain to generate a pure Mach-Zehnder interferometer spectrum. This module can share some parameters in the network structure with the dual-domain mapping module, or it can employ an independent inverse transform algorithm. To ensure the reconstructed spectrum is suitable for valley finding, the spectral reconstruction module can perform amplitude scaling and wavelength coordinate alignment on the output.

[0059] The humidity demodulation module extracts the wavelength shift of the target interference valley in a pure Mach-Zehnder interferometer spectrum and maps the wavelength shift to an ambient humidity value. The module can integrate a multi-valley dynamic routing matrix and a heterogeneous substance diffusion gradient analytical model. Under normal conditions, it directly uses the target interference valley for demodulation; under high humidity and condensation conditions, it switches to a healthy interference valley; and under long-term pollution conditions, it performs refractive index deviation compensation. The final output ambient humidity value can be uploaded to the monitoring platform via a display screen, communication interface, or industrial control bus.

[0060] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for humidity detection of misaligned fusion spliced ​​Mach-Zehnder optical fibers, characterized in that, include: Obtain the original composite interference spectrum output by the misaligned fusion spliced ​​Mach-Zehnder fiber optic humidity sensor; The original composite interference spectrum is input into a pre-trained dual-domain feature mapping network, which converts the original composite interference spectrum from the wavelength domain into a spatial domain equivalent cavity length sequence. Parasitic high-frequency feature peaks with energy density located within a preset isolated interval are extracted from the spatial domain equivalent cavity length sequence. The spatial domain equivalent cavity length sequence and the parasitic high-frequency feature peaks are input into the masking model. The masking model replaces the energy density values ​​corresponding to the parasitic high-frequency feature peaks with zero and outputs a denoised spatial domain sequence. Perform inverse feature mapping on the denoised spatial domain sequence to transform the denoised spatial domain sequence back from the spatial domain to the wavelength domain, generating a pure Mach-Zehnder interferometer spectrum; The wavelength shift of the target interference valley is extracted from the pure Mach-Zehnder interferometer spectrum, and the wavelength shift is mapped to the ambient humidity value.

2. The humidity detection method for misaligned fusion spliced ​​Mach-Zehnder optical fiber according to claim 1, characterized in that, Before inputting the original composite interference spectrum into the pre-trained dual-domain feature mapping network, the steady-state interference spectrum of the misaligned fusion spliced ​​Mach-Zehnder fiber humidity sensor under humidity-free conditions is obtained as sample input data. At the same time, the ideal interference spectrum determined by the interferometer structural parameters is obtained as sample label data. The initial dual-domain feature mapping network is iteratively optimized using the sample input data and sample label data until the mean square error between the output data of the dual-domain feature mapping network and the sample label data converges to within the error convergence condition, thus completing the training of the dual-domain feature mapping network.

3. The humidity detection method for misaligned fusion spliced ​​Mach-Zehnder optical fiber according to claim 1, characterized in that, The steps for extracting parasitic high-frequency feature peaks whose energy density is located within a preset isolated interval from the spatial domain equivalent cavity length sequence include: The spatial domain equivalent cavity length sequence is input into the peak isolation model. The peak isolation model divides the spatial domain equivalent cavity length sequence into a low-frequency energy accumulation region and a high-frequency energy accumulation region, and outputs the isolated peaks that are in the high-frequency energy accumulation region and whose cavity length values ​​are within the preset isolation interval as the parasitic high-frequency characteristic peaks.

4. The humidity detection method for misaligned fusion spliced ​​Mach-Zehnder optical fiber according to claim 1, characterized in that, After generating a pure Mach-Zehnder interferometer spectrum, the peak energy density and central cavity length of the parasitic high-frequency characteristic peaks are extracted. The peak energy density and central cavity length are then input into the aging assessment model. The aging assessment model maps the peak energy density and central cavity length to the coating interface peeling aging index, outputs the coating interface peeling aging index, and triggers an early warning signal.

5. The humidity detection method for misaligned fusion spliced ​​Mach-Zehnder optical fiber according to claim 1, characterized in that, In the step of inputting the spatial domain equivalent cavity length sequence and the parasitic high-frequency feature peak into the masking model, the masking model generates a binarized matrix with the same dimension as the spatial domain equivalent cavity length sequence. The matrix element values ​​of the binarized matrix at the cavity length position corresponding to the parasitic high-frequency feature peak are zero, and the matrix element values ​​of the binarized matrix at the cavity length positions other than the cavity length position corresponding to the parasitic high-frequency feature peak are one. The masking model performs a matrix dot product operation between the binarized matrix and the spatial domain equivalent cavity length sequence to generate the noise-reduced spatial domain sequence.

6. The humidity detection method for misaligned fusion spliced ​​Mach-Zehnder optical fiber according to claim 1, characterized in that, After generating a pure Mach-Zehnder interferometer spectrum, the interference valley sequence corresponding to multiple adjacent cladding modes in the pure Mach-Zehnder interferometer spectrum is obtained. The depth contrast value and center wavelength value of each interference valley in the interference valley sequence are continuously extracted within a predetermined time period to generate a multi-valley contrast time series and a multi-valley center wavelength time series. The multi-valley contrast time series and the multi-valley center wavelength time series are input into the transient droplet scattering identification model. The transient droplet scattering identification model processes the contrast attenuation change of each interference valley through an attention mechanism and outputs the failure probability weight corresponding to each interference valley. The dynamic routing matrix is ​​calculated by taking the failure probability weights of all interference valleys and the time sequence input wavelength of the multi-valley center wavelength. The dynamic routing matrix is ​​used to determine the interference valleys in the time sequence input wavelength of the multi-valley center wavelengths where the failure probability weights have not triggered the routing truncation condition as the target interference valleys.

7. The humidity detection method for misaligned fusion spliced ​​Mach-Zehnder optical fiber according to claim 6, characterized in that, Before inputting the multi-valley contrast time series and multi-valley center wavelength time series into the transient droplet scattering recognition model, the multi-valley contrast time series and multi-valley center wavelength time series under the dehydration phase transition state are collected as positive sample data, and the multi-valley contrast time series and multi-valley center wavelength time series under the stable environment state are collected as negative sample data. The initial transient droplet scattering recognition model is trained using the positive sample data and the negative sample data. During the training process, the attention weight node parameters inside the transient droplet scattering recognition model are adjusted. When the failure probability weight output by the transient droplet scattering recognition model for the positive sample data matches the manually labeled data, the training of the transient droplet scattering recognition model ends.

8. The humidity detection method for misaligned fusion spliced ​​Mach-Zehnder optical fiber according to claim 6, characterized in that, In the step of mapping wavelength offset to ambient humidity value: The wavelength drift trajectory sequence of at least three independent interference valleys of different orders in the pure Mach-Zehnder interferometer spectrum is obtained in multiple time periods. The wavelength drift trajectory sequence is input into the heterogeneous material diffusion gradient analytical model. The heterogeneous material diffusion gradient analytical model extracts the phase difference feature and asynchronous drift rate feature of the independent interference valleys of different orders in the wavelength drift trajectory sequence through a graph neural network. The phase difference feature and asynchronous drift rate feature are mapped to the impurity penetration equivalent refractive index deviation value. The wavelength offset of the target interference valley is subtracted from the equivalent refractive index deviation of the impurity penetration to generate the calibrated wavelength offset, which is then mapped to the ambient humidity value.

9. The humidity detection method for misaligned fusion spliced ​​Mach-Zehnder optical fiber according to claim 8, characterized in that, In the analytical model of diffusion gradient of heterogeneous materials, the graph neural network constructs at least three independent interference valleys as three independent nodes of the graph structure. The graph neural network initializes the edge weights between the three independent nodes to one. During the feature extraction process, the wavelength evolution time lag between the three independent nodes is used as the update compensation term for the edge weights. The refractive index gradual mapping value is calculated using the update compensation term, and the refractive index gradual mapping value is used as the impurity penetration equivalent refractive index deviation value.

10. A humidity detection system for misaligned fusion spliced ​​Mach-Zehnder fiber optic cables for implementing the method of claim 1, characterized in that, The misaligned fusion spliced ​​Mach-Zehnder fiber optic humidity detection system includes: The spectrum acquisition module is used to acquire the original composite interference spectrum output by the misaligned fusion spliced ​​Mach-Zehnder fiber optic humidity sensor. The dual-domain mapping module is used to input the original composite interference spectrum into a pre-trained dual-domain feature mapping network, which converts the original composite interference spectrum from the wavelength domain into a spatial domain equivalent cavity length sequence. The parasitic masking module is used to extract parasitic high-frequency feature peaks with energy density located in a preset isolated interval from the spatial domain equivalent cavity length sequence. The spatial domain equivalent cavity length sequence and the parasitic high-frequency feature peaks are input into the masking model. The masking model replaces the energy density value corresponding to the parasitic high-frequency feature peak with zero and outputs a denoised spatial domain sequence. The spectral reconstruction module is used to perform inverse feature mapping on the denoised spatial domain sequence, converting the denoised spatial domain sequence from the spatial domain back to the wavelength domain, and generating a pure Mach-Zehnder interferometer spectrum. The humidity demodulation module is used to extract the wavelength shift of the target interference valley in the pure Mach-Zehnder interferometer spectrum and map the wavelength shift to the ambient humidity value.

Images