A demodulation system based on sparse spectrum sampling and intelligent feature demodulation cooperation

Abstract

This invention relates to the field of fiber optic grating sensing and spectral measurement technology, and provides a demodulation system based on sparse spectral sampling and intelligent feature demodulation, comprising a demodulator body and a host computer. The demodulator body includes a broadband light source, a beam splitter, a polling measurement unit, a reference unit, a spectral acquisition unit, and a signal processing unit. The broadband light source emits broadband light, which is split into two paths by the beam splitter. The first path from the beam splitter enters an optical switch through a circulator. The reflected light signal from the fiber optic grating sensor returns to the circulator, which transmits the reflected light signal to a coupler. The second path from the beam splitter enters a reference grating. The transmitted light signal from the reference grating is transmitted to the coupler. After the reflected and transmitted light signals are combined in the coupler, the signal processing unit obtains the final spectral data packet to be uploaded to the host computer. This invention enables efficient acquisition and transmission under multi-channel polling conditions, real-time scale calibration, and intelligent demodulation of complex spectral shapes.

Description

Technical Field

[0001] This invention relates to the field of fiber optic grating sensing and spectral measurement technology, and in particular to a demodulation system based on sparse spectral sampling and intelligent feature demodulation. Background Technology

[0002] Fiber Bragg gratings (FBGs) and their arrays are widely used for monitoring physical quantities such as strain, temperature, pressure, and vibration due to their resistance to electromagnetic interference, reusability, and long-distance transmission capabilities. In typical engineering practice, a demodulator often needs to demodulate multiple gratings or even multi-channel arrays simultaneously to reduce system costs and achieve large-scale deployment. On the other hand, tilted fiber Bragg gratings (TFBGs) and SPR-TFBGs formed by coupling with metal films exhibit high sensitivity in refractive index and biochemical detection, but their spectral shapes usually contain dense resonant structures and strong polarization correlations, making demodulation more dependent on high-quality spectra and more complex algorithm processing.

[0003] The closest existing technologies to this application can generally be categorized into three types: The first type is scanning demodulation, which involves rapidly scanning wavelengths using tunable lasers or scanning filters (such as Fabry-Perot, AOTF, etc.), and then obtaining the FBG center wavelength drift through photoelectric detection and calculation. This approach often requires higher scanning frequencies and more complex calibration mechanisms to balance high resolution and high speed. The second type is spectral imaging / spectral demodulation, which uses dispersive elements in conjunction with CCD / CMOS linear arrays to directly obtain spectral sampling sequences, and then uses algorithms to calculate peak positions or features. Spectral demodulators are usually suitable for multi-channel measurements but are constrained by data throughput and processing links, and the corresponding demodulation algorithms are difficult to handle low-quality sparse spectra. The third type is discrete wavelength division / array device demodulation (such as AWG, etc.), which maps wavelength changes to multi-channel intensity allocation, and then inverts the wavelength drift through channel ratios or fitting. This approach can achieve high-frequency measurement and multi-channel parallelism, but this scheme has high requirements for device channel division, calibration, and signal consistency, and the wavelength division multiplexing accuracy is low.

[0004] The aforementioned closest existing technologies have several technical shortcomings in engineering applications: First, under multi-channel polling or large-scale array conditions, if the spectral imaging scheme adopts a full-frame, full-pixel upload approach, it will generate massive amounts of data and significantly increase the communication and computational burden, thereby limiting the measurement frequency or the number of sensors that can be connected. Existing research has also directly pointed out that large-scale FBG arrays are constrained by "massive data volume leading to limited sensor quantity or demodulation frequency" in real-world scenarios, and proposed using FPGAs and other methods to improve real-time processing capabilities. However, the amount of uploaded data is not reduced; only the processing capability is improved. Second, during multi-channel switching / polling, the reference spectrum or calibration information often requires additional measurement or cannot be continuously visible in each sampling, causing temperature drift, light source drift, and device drift to accumulate on the time axis as baseline inconsistency and scale drift, thus affecting long-term stability. Some works improve stability through reference gratings or adaptive calibration, but there are still structural contradictions between polling, real-time reference persistence, and low-data-volume upload. Third, in complex spectral scenarios such as TFBG / SPR-TFBG, the dense spectral lines, polarization sensitivity, and sensitivity to environmental disturbances cause scanning and array demodulation systems to fail. Traditional full-spectrum peak finding / fitting is more prone to misjudgment or decoupling failure when there is sparse sampling, low signal-to-noise ratio, or spectral distortion. Related studies show that demodulation of TFBG / SPR-TFBG often requires more complex processing (such as polarization / phase analysis) or higher quality data acquisition conditions, thus creating new engineering bottlenecks in terms of cost, real-time performance, and robustness.

[0005] Therefore, there is an urgent need for a demodulation system that is oriented towards sparse sampling, multi-channel polling, and complex spectral demodulation, so as to overcome the shortcomings of the existing technologies in terms of throughput, stability, and industrial deployment of complex spectral demodulation. Summary of the Invention

[0006] This invention primarily addresses the long-standing technical challenges of existing fiber Bragg grating demodulation technologies in engineering applications, namely, difficulties in multi-channel high-frequency measurement, long-term stability, and usability for complex spectral shapes. It proposes a demodulation system based on sparse spectral sampling and intelligent feature demodulation, achieving efficient acquisition and transmission under multi-channel polling conditions, real-time scale calibration, intelligent demodulation of complex spectral shapes, and safe fallback under abnormal operating conditions, without relying on large-scale full-spectrum data streams and frequent manual calibration. This results in a long-term, scalable, and integrated hardware and software fiber Bragg grating demodulation system.

[0007] This invention provides a demodulation system based on sparse spectral sampling and intelligent feature demodulation, comprising: a demodulator body and a host computer;

[0008] The demodulator body includes: a broadband light source, a beam splitter, a polling measurement unit, a reference unit, a spectrum acquisition unit, and a signal processing unit;

[0009] The broadband light source is used to emit broadband light covering the target measurement band. After being split into two paths by a beam splitter, the first path of light from the beam splitter enters an optical switch through a circulator. The optical switch is used to connect multiple fiber Bragg grating sensors of the same or different types. The reflected light signal from the fiber Bragg grating sensor returns to the circulator, which transmits the reflected light signal to a coupler.

[0010] The second beam from the beam splitter enters the reference grating; the transmitted light signal from the reference grating is transmitted to the coupler.

[0011] After the reflected and transmitted light signals are combined in the coupler, the combined signal is acquired by the spectral acquisition unit; the combined signal includes the linear array pixel index and the spectral intensity sequence.

[0012] The signal processing unit adopts a second sparse upload strategy, which means processing the combined signal to obtain the final spectral data packet to be uploaded to the host computer.

[0013] Preferably, the polling measurement unit includes a circulator, an optical switch, and multiple fiber Bragg grating sensors of the same or different types.

[0014] Preferably, the fiber Bragg grating sensor includes, but is not limited to, fiber Bragg gratings, tilted fiber Bragg gratings, phase-shifting gratings, and long-period fiber Bragg gratings.

[0015] Preferably, the spectral acquisition unit adopts a first sparse sampling strategy;

[0016] The first sparse sampling strategy includes: the spectral sampling interval is greater than or equal to The order of magnitude is in the order of pm, and the total number of pixels is less than or equal to 1000. Magnitude.

[0017] Preferably, the second sparse uploading strategy includes: dividing the bundled signal into four functional partitions and determining the connected window range corresponding to each functional partition;

[0018] The four functional zones are respectively the noise floor region set. Reference Channel Area Collection TFBG slow zone collection With FBG fast zone collection The set of each functional partition includes the linear array pixel index sequence and the corresponding spectral intensity sequence under a specific spectral range;

[0019] After each frame is acquired, the signal processing unit performs initial screening based on the light intensity threshold within the combined signal to obtain a linear array pixel index set. The threshold is determined according to the attenuation degree of each channel, and then functional partitions are divided based on the threshold.

[0020] The set of linear pixel indices for each functional partition can be defined as:

[0021] ;

[0022] Where k represents the name of each partition; Represents the set of linear pixel indices for each partition; This represents a set of functional zones, selected from the set of noise floor regions. Reference Channel Area Collection TFBG slow zone collection With FBG fast zone collection ; Indicates the linear array pixel index; Represents a spectral intensity sequence; This indicates the threshold value for the corresponding partition;

[0023] The signal processing unit classifies the linear array pixel indexes in the order of FBG fast region, TFBG slow region, reference channel region, and noise base region. When the spectral intensity under index i is greater than the threshold of the corresponding partition, the index is classified into the corresponding partition.

[0024] The signal processing unit determines the connectivity range of each signal partition based on the divided signal partitions, and finally packages and uploads the spectral data packets to the host computer.

[0025] The final expression for uploading the spectral data packet is:

[0026] ;

[0027] in This indicates the spectral data packet that the information processing unit finally uploads to the host computer, including the pixel index and intensity sequence under the connectivity range of each partition; For reference, the connectivity range of the channel area; This refers to the connectivity range of the FBG fast zone; The connectivity range of the TFBG slow region; Select switch; This refers to the noise floor region.

[0028] Preferably, the host computer receives spectral data packets from the signal processing unit; the host computer performs preprocessing based on the noise floor region and the reference channel region to obtain calibrated measurement spectral information;

[0029] Preprocessing includes baseline processing and real-time calibration, which are performed using the following formulas:

[0030] ;

[0031] in, This indicates the measurement spectrum information for the corresponding serial number i after calibration. The reference spectrum amplitude matching coefficients, To upload spectral intensity, This indicates the base strength estimate. This indicates the intensity estimate of the reference spectrum.

[0032] Preferably, the host computer performs a spectral complexity determination on the calibrated measurement spectrum to determine whether the calibrated measurement spectrum information is single-peaked or multi-peaked;

[0033] The spectral complexity is defined as the normalized total variation:

[0034] ;

[0035] C represents the normalized total variation; For the calibrated measurement spectrum information, The threshold is a constant; then the data enters the gating unit, which selects the demodulation path according to the decision rule; the decision rule is a specified threshold. ,when When the measured spectrum information is considered as a single peak, when C is greater than or equal to the threshold, it is judged as a multi-peak.

[0036] Subsequently, the data is determined according to the complexity of the spectral shape and enters a specific demodulation process under gating control; if the measured spectrum is single-peaked, the host computer uses peak finding fitting or frequency domain processing to demodulate the feature quantities; if the measured spectrum is multi-peaked, the host computer uses deep learning algorithms to demodulate the feature quantities.

[0037] Preferably, the deep learning algorithm includes the following processes:

[0038] The TFBG input is represented as a sparse vector after partition selection. Input to pre-trained neural network Complete multi-parameter regression and decoupling:

[0039] ;

[0040] in, For model parameters, This is the output of the multi-parameter function to be demodulated.

[0041] Preferably, the host computer synchronously calculates and evaluates the out-of-range fraction of the intelligent demodulation output based on the Mahalanobis distance distribution in the feature space:

[0042] ;

[0043] in, Represents out-of-distribution fractions. The feature vector of the model at a specified layer; For the first Class / First The feature mean of each working condition training set; For Mahalanobis metric matrix; when Trigger rollback and output a warning when; The system accepts multiple parameters from the current neural network model output.

[0044] Preferably, the deep learning model used for demodulating complex spectral data is obtained and updated through local transfer learning;

[0045] The local transfer learning includes: the host computer collecting standard operating condition data locally. In the original parameters of the neural network Based on this, new parameters of the neural network are obtained through transfer learning. Its local adaptation process can be abstracted as follows:

[0046] ;

[0047] in, This represents the training objective on local data. To maintain consistency with the pre-trained model, regularization constraints For weights.

[0048] The present invention provides a demodulation system based on sparse spectral sampling and intelligent feature demodulation, which has the following advantages compared with the prior art:

[0049] 1. The advantages of this application mainly come from the four-zone window upload based on light intensity threshold at the acquisition end, the resident calibration of the reference spectrum, the polling scheduling of fast and slow channels, and the gating and rollback mechanism of the host computer. These measures respectively improve data link efficiency, long-term stability, and security under abnormal operating conditions.

[0050] 2. Regarding work efficiency, the FPGA side no longer uploads the entire frame and all pixels, but only uploads consecutive valid windows that meet the threshold conditions in the four partitions. This makes the amount of uploaded data change with the valid information rather than with the full spectrum length, thereby reducing communication bandwidth and the processing burden on the host computer, and improving the effective refresh rate under multi-channel operation. A single channel of the device can meet the stable operation of 3kHz, with an overall polling and demodulation rate of 1kHz. Under full-speed operation, the FBG fast zone can achieve detection of frequencies up to 3kHz, while the strain measurement channel maintains a micro-strain accuracy of 5.

[0051] 3. In terms of reliability, the reference channel is continuously visible in each frame of spectrum in a superimposed manner, which provides a real-time calibration basis for temperature drift, light source drift and device drift, avoiding calibration gaps caused by the invisibility of the reference under the polling structure, thereby improving long-term stability and repeatability, and the whole can run stably at full speed for more than 5 hours.

[0052] 4. In terms of security and engineering availability, the host computer adopts gating and diversion: simple spectral patterns are prioritized to use the traditional fast demodulation link, while complex TFBG spectral patterns enter intelligent demodulation; when the confidence level or OOD index exceeds the threshold, it triggers a fallback to traditional demodulation or conservative output. If necessary, it can also link with the acquisition side to adjust the slow-speed sampling strategy to avoid misjudgment and loss of lock under abnormal working conditions; thus realizing intelligent feature demodulation.

[0053] 5. Regarding adaptability, the TFBG model utilizes local transfer learning to quickly adapt to different devices / tasks with limited calibration data, lowering the deployment and maintenance threshold. The refractive index error before and after transfer is within [a certain range]. It operates at the RIU level without affecting the fast region demodulation of FBG.

[0054] 6. This invention can be applied to structural health monitoring, strain / temperature / vibration monitoring of rail transit and industrial equipment, real-time measurement of multi-point multi-channel fiber optic sensing arrays, and refractive index / biochemical sensing and multi-parameter measurement based on complex spectral shapes such as tilted fiber grating (TFBG) / plasmocouple (SPR). Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the composition of each unit and the signal transmission process of the demodulation system based on sparse spectral sampling and intelligent feature demodulation collaboration provided by the present invention.

[0056] Figure 2 This is a functional partitioning diagram of the information processing unit for the combined signal under the FBG fast zone polling channel;

[0057] Figure 3 This is a functional partitioning diagram of the information processing unit for the combined signal under the TFBG slow zone polling channel;

[0058] Figure 4 This is a diagram showing the effect of the host computer simultaneously demodulating TFBG and FBG data packets;

[0059] Figure 5 This is a flowchart of the host computer demodulation algorithm. Detailed Implementation

[0060] To make the technical problems solved by this invention, the technical solutions adopted, and the technical effects achieved clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings, not all of them.

[0061] like Figure 1As shown in the figure, an embodiment of the present invention provides a demodulation system based on sparse spectral sampling and intelligent feature demodulation collaboration, comprising: a demodulator body and a host computer;

[0062] The demodulator body includes: a broadband light source, a beam splitter, a polling measurement unit, a reference unit, a spectrum acquisition unit, and a signal processing unit.

[0063] The polling measurement unit includes a circulator, an optical switch, and multiple fiber Bragg grating sensors of the same or different types. Fiber Bragg grating sensors include, but are not limited to, fiber Bragg gratings (FBGs), tilted fiber Bragg gratings (TFBGs), phase-shifted fiber Bragg gratings (PSFBGs), and long-period fiber gratings (LPFGs). Any fiber Bragg grating that can be demodulated by intensity or wavelength shift can be connected to the polling measurement unit. This unit contains multiple optical paths, with each or more gratings located on one path. Each path can be further divided into a fast FBG region and a slow TFBG region.

[0064] The polling measurement unit uses an optical switch to achieve multi-channel time-division multiplexing, so as to access multiple fiber Bragg grating sensors of the same or different types.

[0065] The reference unit includes a reference optical fiber. The reference optical fiber is preferably a temperature scale grating or a reference FBG with a known temperature coefficient. The reference optical fiber used employs reflection spectrum demodulation, and the reference unit may require a circulator to receive the reflection spectrum. The reference unit provides traceable, real-time calibration information.

[0066] The broadband light source is used to emit broadband light covering the target measurement band. After being split into two paths by a beam splitter, the first path of light from the beam splitter enters an optical switch through a circulator. The optical switch is used to connect multiple fiber Bragg grating sensors of the same or different types. The reflected light signal from the fiber Bragg grating sensor returns to the circulator, which transmits the reflected light signal to a coupler. The broadband light is preferably an amplified spontaneous emission light source, a superluminescent diode light source, or other light source capable of outputting broadband light.

[0067] The second beam from the beam splitter enters the reference grating; the transmitted light signal from the reference grating is transmitted to the coupler.

[0068] After the reflected and transmitted light signals are combined in the coupler, the combined signal is acquired by the spectral acquisition unit.

[0069] In this invention, the two optical paths are combined before entering the same spectral acquisition unit, so that each frame of the combined signal contains both the measurement spectrum of the polling measurement unit and the reference spectrum of the reference unit, realizing a superposition sampling mechanism of "reference spectrum is always present, polling spectrum is switched".

[0070] The spectral acquisition unit is used to perform spectral decomposition on the combined optical signal and acquire discretely sampled spectral data. Preferably, it is a dispersive device and a CMOS / CCD linear or area array. The acquired combined signal includes the linear array pixel index and the spectral intensity sequence, covering the target wavelength band with a finite number of pixels and outputting the discretely sampled spectral intensity sequence. The linear array pixel index of the combined signal is i = 1, 2, ..., N, corresponding to a wavelength mapping of λ(i), and the spectral intensity sequence of the acquired combined signal is represented as I(i). As an optional implementation example, the number of pixels in the linear array detector can be 256 or 512, corresponding to coverage of approximately 80 nm wavelength range.

[0071] The spectral acquisition unit of this invention employs a first-stage sparse sampling strategy in the pixel count implementation. In this invention, the spectral sampling interval and the total number of pixels corresponding to adjacent pixels differ from traditional schemes, compared to the optical spectrum analyzer (OSA) and optical vector analyzer (OVA) used in traditional fiber optic demodulation systems. The first-stage sparse sampling strategy includes: a spectral sampling interval greater than or equal to... The order of magnitude is in the order of pm, and the total number of pixels is less than or equal to 1000. The spectral sampling interval of this invention is increased from the order of 10 pm to... The number of pixels is in the pm range. The magnitude dropped to This significantly reduces the amount of data. In the implementation example, the total number of pixels in the spectral signal is 512. An OSA demodulation system in the same wavelength range would require more than 5,000 pixels (this value is an example to illustrate the achievable sampling scale and beneficial effects, and does not constitute the sole limitation on device specifications), thus providing the hardware foundation for subsequent windowed acquisition and fast demodulation.

[0072] The key to this invention lies in performing a first sparse acquisition in the spectral acquisition unit and a second sparse upload in the signal processing unit. These two processes, along with the reference superposition mechanism, fast and slow channel strategy, and host computer gated demodulation, form a closed-loop collaboration. The hardware-side signal processing unit no longer uses a full-frame, full-pixel upload method. Instead, it performs windowed selection of the superimposed spectrum based on a light intensity threshold and preset partition boundaries, uploading only the set of pixels carrying valid information. This reduces input and output bandwidth usage and the processing burden on the host computer from the source.

[0073] The signal processing unit adopts a second sparse upload strategy, which means processing the combined signal to obtain the final spectral data packet to be uploaded to the host computer.

[0074] The second sparse upload strategy includes: dividing the combined signal into four functional zones, determining the connected window range corresponding to each functional zone, and finally packaging the spectral data packet and uploading it to the host computer.

[0075] The signal processing unit divides the combined signal into four functional zones.

[0076] like Figure 2 and Figure 3 As shown, the four functional zones are sets of noise floor regions. Reference Channel Area Collection TFBG slow zone collection With FBG fast zone collection The spectral signals from the FBG fast region and the reference channel region are uploaded to the host computer after the connectivity window expansion range is further determined. Since the polling channel is the TFBG slow region, the combined signal does not contain the signal from the FBG fast region. At this time, the signal processing unit further processes the spectral signals from the TFBG fast region and the reference channel region, and after the connectivity window expansion range is determined, they are uploaded to the host computer.

[0077] Each functional zone set includes a linear array pixel index sequence and a corresponding spectral intensity sequence within a specific spectral range. The reference channel region set... Used to provide a continuously visible calibration scale per frame; FBG fast area collection Narrowband sampling around the FBG peaks is used to support high refresh rates; TFBG slow region set Used to cover the effective spectral band of TFBG to support demodulation of complex spectral shapes; noise floor region set Used for dark noise characterization and threshold benchmark determination, but not required for complete upload of each frame.

[0078] After each frame is acquired, the signal processing unit first performs initial screening within the combined signal based on the light intensity threshold to obtain a linear array pixel index set. Then, it determines the threshold based on the attenuation degree of each channel and divides the functional areas based on the threshold.

[0079] The set of linear pixel indices for each functional partition can be defined as:

[0080] ;

[0081] Where k represents the name of each partition; Represents the set of linear pixel indices for each partition; This represents a set of functional zones, selected from the set of noise floor regions. Reference Channel Area Collection TFBG slow zone collection With FBG fast zone collection ; Indicates the linear array pixel index; Represents a spectral intensity sequence; This indicates the threshold for the corresponding partition.

[0082] The signal processing unit classifies the linear array pixel indexes in the order of FBG fast region, TFBG slow region, reference channel region, and noise base region. When the spectral intensity under index i is greater than the threshold of the corresponding partition, the index is classified into the corresponding partition.

[0083] The signal processing unit determines the connectivity range of each signal partition based on the divided signal partitions, and finally packages and uploads the spectral data packets to the host computer.

[0084] To avoid missing meaningful rising and falling edges or introducing scattered invalid points by only uploading peak points, the signal processing unit further sets the linear array pixel index set at time t. The main connected regions are selected and appropriately expanded (i.e., elements are added to the linear pixel index set) to obtain the final connected regions. This ensures that the uploaded measurement spectrum is continuous along the pixel axis, can be used for subsequent demodulation, and does not introduce fragment points unrelated to the partition.

[0085] Reference channel range As real-time calibration information, it is uploaded in every frame; whether the FBG fast region and TFBG slow region are uploaded is determined by the polling channel type and the fast / slow scheduling switch, and the corresponding final uploaded spectral data packet expression is:

[0086] ;

[0087] in This indicates the spectral data packet that the information processing unit finally uploads to the host computer, including the pixel index and intensity sequence under the connectivity range of each partition; For reference, the connectivity range of the channel area; This refers to the connectivity range of the FBG fast zone; The connectivity range of the TFBG slow region; The selection switch is used because FBG and TFBG will not be located on the same channel. Therefore, at any given time, only one of them will be in the combined signal. The selection switch determines the current signal type based on the optical switch channel. When the switch is 0, the corresponding item is an empty set, that is, there is no signal of that type, and there is no need to upload the connectivity range of the corresponding partition. For the noise floor region, it is preferable to upload only in dark noise calibration frames or sampling monitoring mode to support threshold determination and baseline estimation; otherwise, it should not be uploaded to reduce data sparsity. Reference channel region Carrying reference grating spectral information, it remains continuously visible in each frame, thus providing a real-time scale for temperature or drift calibration, with separate reference and measurement channel areas. If the spectral domain separability condition is met, and the host computer detects a conflict between the two, frame extraction monitoring can be performed; TFBG slow partitioning. Allows for a relatively wider window and a relatively larger number of sampling points, but acquires them at a lower refresh rate to match their relatively slow changing characteristics; FBG fast partitioning. A narrow window is set around the peak of FBG to achieve high refresh rate output with fewer sampling points, which matches the rapid changes in FBG during dynamic measurement.

[0088] The second sparse upload strategy includes the above two steps: the first step is initial screening, where the signal processing unit divides the signal into functional zones based on a preset light intensity threshold after each frame is acquired; the second step is packaging, where the connectivity range of each zone is filtered and expanded, and finally, the information is selectively packaged into spectral data packets for upload, with some data discarded to achieve the effect of sparse data volume. The threshold value, the number of connectivity range expansion points, and the width of each zone range can be determined according to the device calibration results or operating mode configuration to adapt to different spectral ranges, different sensor types, and different application conditions; this application does not limit specific values, in order to maintain feasibility while reserving sufficient space for subsequent product iterations.

[0089] The signal processing unit also includes: controlling the optical switch to cooperate with the second sparse upload strategy function, that is, controlling the polled channel at a certain time and the time spent collecting data in that channel.

[0090] Specifically, regarding multi-channel polling, this application divides the polling channels into high-speed and low-speed subsets, and achieves "fast-slow decoupling" polling scheduling through a data acquisition-side partitioning mechanism and a host computer strategy. The high-speed subset preferably corresponds to FBG-type sensor channels, combined with... Figure 2 The FBG fast region narrow window and high frame rate acquisition shown enable high refresh rate and low data volume fast output; the low speed subset is selected to correspond to the TFBG type sensor channel, combined with the TFBG slow region wide window and lower acquisition frequency, to achieve full characterization of complex spectral patterns but avoid causing continuous high load on the optical switch and data link.

[0091] The central control unit can adjust the optical switch dwell time, the ratio of fast and slow channels, and the acquisition frequency of the slow zone according to the task parameters issued by the host computer, so that the system can maintain stable throughput and controllable latency when running in multi-channel mode.

[0092] Furthermore, since the reference channel area is preserved in each frame, the polling process no longer needs to occupy an additional polling time slot to sample the reference spectrum, thereby avoiding calibration gaps caused by the invisibility of the reference and reducing the risk of scale inconsistency caused by changes in the polling clock.

[0093] like Figure 5 As shown, in the host computer, the present invention adopts a unified process of "gating and diversion - traditional demodulation - intelligent demodulation - confidence backoff".

[0094] The host computer receives spectral data packets from the signal processing unit.

[0095] The host computer performs preprocessing based on the noise floor region and the reference channel region to obtain the calibrated measurement spectrum information.

[0096] Preprocessing includes baseline processing and real-time calibration. The calibrated measurement spectrum information is obtained through basis subtraction and reference spectrum extraction, making subsequent demodulation less sensitive to light source fluctuations, device drift and environmental temperature drift.

[0097] Baseline processing and real-time calibration are performed using the following formula:

[0098] ;

[0099] in, This indicates the measurement spectrum information for the corresponding serial number i after calibration. The reference spectrum amplitude matching coefficients, To upload spectral intensity, This represents the base strength estimate (which can be obtained statistically from the noisy base region). This represents the reference spectral intensity estimate (which can be extracted from the reference channel region and mapped to the global index).

[0100] The host computer then performs a spectral complexity determination on the calibrated measurement spectrum to determine whether the calibrated measurement spectrum information is single-peaked or multi-peaked. In one implementation example, the spectral complexity can be defined as the normalized total variation:

[0101] ;

[0102] C represents the normalized total variation; For the calibrated measurement spectrum information, This is a constant. The data then enters the gating unit, which selects the demodulation path according to a decision rule. The decision rule is a specified threshold. ,when When measuring spectral information, it is considered a single peak; when C is greater than or equal to the threshold, it is judged as multi-peak. This application is not limited to the above judgment; in practice, it can be modified to determine spectral complexity by measuring the number of peaks or the half-width ratio, etc., as needed.

[0103] Subsequently, the data is determined by spectral complexity and gated into a specific demodulation process. For single-peaked spectra, the host computer uses traditional algorithms such as peak finding and fitting or frequency domain processing for feature demodulation. For simple spectral data corresponding to the fast region of the FBG, the system can directly enter a traditional fast demodulation path (e.g., peak finding and fitting or frequency domain processing) to obtain a high refresh rate output with minimal computational overhead. This algorithm branch outputs parameters after quickly completing demodulation and reference calibration. If the measured spectrum is multi-peaked, the host computer uses a deep learning algorithm for feature demodulation.

[0104] For the complex spectral data corresponding to the slow region of TFBG, the deep learning algorithm used by the system includes the following process:

[0105] The TFBG input is represented as a sparse vector after partition selection. Input to pre-trained neural network Complete multi-parameter regression and decoupling:

[0106] ;

[0107] in, For model parameters, This is the multi-parameter output to be demodulated. The neural network form, loss structure, and training details here do not constitute a limitation on the scope of protection of this application. They can be selected according to different sensors and different task requirements, thereby preserving the scalability of the technical approach.

[0108] To ensure project safety, the host computer synchronously calculates the out-of-distribution (OOD) score based on the Mahalanobis distance in the feature space for the intelligent demodulation output, and performs evaluation based on this OOD score:

[0109] ;

[0110] The out-of-distribution score is obtained by training with spectral data collected under specific categories or operating conditions, where Represents out-of-distribution (OOD) scores. The feature vector of the model at a specified layer; For the first Class / First The feature mean of each working condition training set; This is a Mahalanobis metric matrix. When... When a rollback is triggered and a warning is output, the system can switch to a traditional demodulation algorithm to demodulate the reconstructed spectrum or available spectral bands, or output a conservative estimate. It can also choose to send adjustment commands to the acquisition end (expanding the window or reducing the polling frequency) to avoid misjudgments and lockouts under abnormal operating conditions, achieving a demodulation strategy that combines performance improvement and safety fallback. The system accepts multiple parameters from the current neural network model. The algorithm branch outputs parameters after the above OOD evaluation.

[0111] Considering the difficulty in model generalization caused by differences between TFBG devices, the deep learning model used for demodulation of complex spectral data in this invention is obtained and updated through local transfer learning.

[0112] The local transfer learning refers to the following: the host computer collects standard operating condition data locally. In the original parameters of the neural network Based on this, new parameters of the neural network are obtained through transfer learning. Its local adaptation process can be abstracted as follows:

[0113] ;

[0114] in, This represents the training objective on local data. To maintain consistency with the pre-trained model, regularization constraints The weights are used as parameters. Transfer learning strategies include freezing some feature layers and updating the regression head, fine-tuning only a few layer parameters, or employing consistency constraints, or a combination thereof. This application does not limit the specific training strategy and parameter details to avoid limiting the scope of protection to a single implementation. Through local transfer learning, the system can quickly deploy different TFBGs at a lower calibration cost and can be iteratively updated during equipment maintenance or changes in operating conditions, improving long-term availability.

[0115] Through the above technical solutions, this application achieves advantages that previous technologies do not possess in terms of system throughput, stability, and collaborative demodulation of FBG and TFBG spectral shapes: the two sparse mechanisms performed at the demodulator end make the amount of uploaded data and the computational load of the host computer increase with the effective information rather than the full spectrum length, thereby significantly reducing the data link burden and shortening the end-to-end latency under multi-channel polling conditions; the superimposed reference channel area is continuously visible in each frame, and combined with the host computer calibration process, it makes the system more stable against temperature drift, device drift, and light source fluctuations, reducing calibration gaps under polling structures; the gating and offloading and OOD backoff mechanisms in the software unit enable intelligent demodulation of complex spectral shapes under sparse sampling conditions, while having a controllable safe degradation path under abnormal operating conditions; the local transfer learning mechanism enables the TFBG model to quickly adapt to different devices with a small amount of local calibration data, lowering the threshold for industrial deployment.

[0116] The overall beneficial effects of this invention are evident. Figure 4 For data packets in the fast region of FBG, the program demodulates the vibrational frequency; for data packets in the slow region of TFBG, the program demodulates the environmental refractive index. These two processes occur simultaneously. When both FBG and TFBG branches exist in the polling channel, the system demodulates both spectra simultaneously. It is evident that demodulating the vibrational frequency in the fast region of FBG, compared to strain gauges, demonstrates higher overall accuracy and fewer glitches. Demoting the standard liquid refractive index in the slow region of TFBG, deep learning can stably output the results under sparse sampling. Furthermore, both are polled measurements and do not interfere with each other.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions for some or all of the technical features, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A demodulation system based on sparse spectral sampling and intelligent feature demodulation, characterized in that, include: Demodulator unit and host computer; The demodulator body includes: a broadband light source, a beam splitter, a polling measurement unit, a reference unit, a spectrum acquisition unit, and a signal processing unit; The broadband light source is used to emit broadband light covering the target measurement band. After being split into two paths by a beam splitter, the first path of light from the beam splitter enters an optical switch through a circulator. The optical switch is used to connect multiple fiber Bragg grating sensors of the same or different types. The reflected light signal from the fiber Bragg grating sensor returns to the circulator, which transmits the reflected light signal to a coupler. The second beam from the beam splitter enters the reference grating; the transmitted light signal from the reference grating is transmitted to the coupler. After the reflected and transmitted light signals are combined in the coupler, the combined signal is acquired by the spectral acquisition unit; the combined signal includes the linear array pixel index and the spectral intensity sequence. The signal processing unit adopts a second sparse upload strategy, which means processing the combined signal to obtain the final spectral data packet to be uploaded to the host computer.

2. The demodulation system based on sparse spectral sampling and intelligent feature demodulation coordination according to claim 1, characterized in that, The polling measurement unit includes a circulator, an optical switch, and multiple fiber optic grating sensors of the same or different types.

3. The demodulation system based on sparse spectral sampling and intelligent feature demodulation coordination according to claim 1, characterized in that, The fiber Bragg grating sensor includes, but is not limited to, fiber Bragg gratings, tilted fiber Bragg gratings, phase-shifting gratings, and long-period fiber Bragg gratings.

4. The demodulation system based on sparse spectral sampling and intelligent feature demodulation coordination according to claim 1, characterized in that, The spectral acquisition unit employs a first-stage sparse sampling strategy; The first sparse sampling strategy includes: the spectral sampling interval is greater than or equal to The order of magnitude is in the order of pm, and the total number of pixels is less than or equal to 1000. Magnitude.

5. The demodulation system based on sparse spectral sampling and intelligent feature demodulation coordination according to claim 1, characterized in that, The second sparse upload strategy includes: dividing the bundled signal into four functional partitions and determining the connected window range corresponding to each functional partition; The four functional zones are respectively the noise floor region set. Reference Channel Area Collection TFBG slow zone collection With FBG fast zone collection The set of each functional partition includes the linear array pixel index sequence and the corresponding spectral intensity sequence under a specific spectral range; After each frame is acquired, the signal processing unit performs initial screening based on the light intensity threshold within the combined signal to obtain a linear array pixel index set. The threshold is determined according to the attenuation degree of each channel, and then functional partitions are divided based on the threshold. The set of linear pixel indices for each functional partition can be defined as: ； Where k represents the name of each partition; Represents the set of linear pixel indices for each partition; This represents a set of functional zones, selected from the set of noise floor regions. Reference Channel Area Collection TFBG slow zone collection With FBG fast zone collection ; Indicates the linear array pixel index; Represents a spectral intensity sequence; This indicates the threshold value for the corresponding partition; The signal processing unit classifies the linear array pixel indexes in the order of FBG fast region, TFBG slow region, reference channel region, and noise base region. When the spectral intensity under index i is greater than the threshold of the corresponding partition, the index is classified into the corresponding partition. The signal processing unit determines the connectivity range of each signal partition based on the divided signal partitions, and finally packages and uploads the spectral data packets to the host computer. The final expression for uploading the spectral data packet is: ； in This indicates the spectral data packet that the information processing unit finally uploads to the host computer, including the pixel index and intensity sequence under the connectivity range of each partition; For reference, the connectivity range of the channel area; This refers to the connectivity range of the FBG fast zone; The connectivity range of the TFBG slow region; Select switch; This refers to the noise floor region.

6. The demodulation system based on sparse spectral sampling and intelligent feature demodulation coordination according to claim 1, characterized in that, The host computer receives spectral data packets from the signal processing unit; the host computer performs preprocessing based on the noise floor region and the reference channel region to obtain calibrated measurement spectrum information; Preprocessing includes baseline processing and real-time calibration, which are performed using the following formulas: ； in, This indicates the measurement spectrum information for the corresponding serial number i after calibration. The reference spectrum amplitude matching coefficients, To upload spectral intensity, This indicates the base strength estimate. This indicates the intensity estimate of the reference spectrum.

7. The demodulation system based on sparse spectral sampling and intelligent feature demodulation coordination according to claim 6, characterized in that, The host computer performs a spectral complexity assessment on the calibrated measurement spectrum to determine whether the calibrated measurement spectrum information is single-peaked or multi-peaked. The spectral complexity is defined as the normalized total variation: ； C represents the normalized total variation; For the calibrated measurement spectrum information, The threshold is a constant; then the data enters the gating unit, which selects the demodulation path according to the decision rule; the decision rule is a specified threshold. ,when When the measured spectrum information is considered as a single peak, when C is greater than or equal to the threshold, it is judged as a multi-peak. Subsequently, the data is determined according to the complexity of the spectral shape and enters a specific demodulation process under gating control; if the measured spectrum is single-peaked, the host computer uses peak finding fitting or frequency domain processing to demodulate the feature quantities; if the measured spectrum is multi-peaked, the host computer uses deep learning algorithms to demodulate the feature quantities.

8. The demodulation system based on sparse spectral sampling and intelligent feature demodulation coordination according to claim 7, characterized in that, The deep learning algorithm includes the following processes: The TFBG input is represented as a sparse vector after partition selection. Input to pre-trained neural network Complete multi-parameter regression and decoupling: ； in, For model parameters, This is the output of the multi-parameter function to be demodulated.

9. The demodulation system based on sparse spectral sampling and intelligent feature demodulation coordination according to claim 8, characterized in that, The host computer evaluates the out-of-range fraction of the intelligent demodulation output based on the Mahalanobis distance distribution in the feature space: ； in, Represents out-of-distribution fractions. The feature vector of the model at a specified layer; For the first Class / First The feature mean of each working condition training set; For Mahalanobis metric matrix; when Trigger rollback and output a warning when; The system accepts multiple parameters from the current neural network model output.

10. The demodulation system based on sparse spectral sampling and intelligent feature demodulation coordination according to claim 9, characterized in that, The deep learning model used for demodulating complex spectral data is obtained and updated through local transfer learning; The local transfer learning includes: the host computer collecting standard operating condition data locally. In the original parameters of the neural network Based on this, new parameters of the neural network are obtained through transfer learning. Its local adaptation process can be abstracted as follows: ； in, This represents the training objective on local data. To maintain consistency with the pre-trained model, regularization constraints For weights.

Images