Machine learning based pgc light intensity and phase mapping demodulation method and apparatus

Abstract

The application relates to a PGC light intensity and phase mapping demodulation method and device based on machine learning. The method comprises the following steps: a single-frequency laser is injected into an interference type optical fiber sensor, a data acquisition card is used to apply a PGC high-frequency carrier through internal / external modulation, and interference light carrying a to-be-measured physical quantity is output by the sensor; an optoelectronic detector is used to convert the interference light into an electrical signal, and a normalized long-time-domain light intensity digital signal is obtained through the data acquisition card. The demodulation device divides the signal according to a single PGC cycle, performs Fourier transform, inputs frequency domain features into a random forest regression model, outputs a modulation depth, and feeds back a stable modulation voltage. The stable signal is input into a support vector machine regression model optimized by a particle swarm, an initial phase is solved, splicing is performed, direct current is eliminated, and a phase time delay is eliminated, and finally, a to-be-measured physical quantity is demodulated, so that the system demodulation cost can be reduced.

Description

Technical Field

[0001] This application relates to the field of interferometric fiber optic sensing technology, and in particular to a PGC light intensity and phase mapping demodulation method and apparatus based on machine learning. Background Technology

[0002] Phase-generated carrier (PGC) is a conventional demodulation method for interferometric phase-change fiber optic sensors. In interferometric fiber optic sensing systems, the traditional PGC method first modulates the single-frequency light output from a single-frequency laser using a high-frequency carrier through external or internal modulation. This modulated single-frequency light is then input into the interferometric fiber optic sensor, which outputs light carrying the target signal. The output light is processed by a photodetector and a data acquisition card to generate a digital signal, which is then sent to a signal processing demodulation device. Finally, the target signal is demodulated through carrier signal mixing, low-pass filtering, arctangent, or differential multiplication. The traditional PGC demodulation process is complex, especially in large-scale interferometric fiber optic sensor array systems, requiring significant computational resources and resulting in high power consumption. Furthermore, during operation, the PGC demodulation parameters of the interferometric fiber optic sensing system often change due to variations in environmental parameters. Since the PGC demodulation results are sensitive to these parameters, rapid and accurate calculation of these parameters is crucial for ensuring the accuracy of the demodulation results. However, existing PGC demodulation parameter calculation methods suffer from high computational complexity and other drawbacks. By overcoming the existing limitations of traditional PGC demodulation methods, the sensing performance of interferometric fiber optic sensing systems can be improved. Summary of the Invention

[0003] Therefore, it is necessary to provide a machine learning-based PGC intensity and phase mapping demodulation method and apparatus that can reduce the demodulation overhead of interferometric fiber optic sensing systems, in order to address the aforementioned technical problems.

[0004] A machine learning-based PGC intensity and phase mapping demodulation method is applied to an interferometric fiber optic sensing system consisting of a single-frequency laser, an interferometric fiber optic sensor, a photodetector, a data acquisition card, and a signal processing and demodulation device connected in sequence. The signal processing and demodulation device incorporates a random forest regression model, a particle swarm optimization support vector machine regression model, and a feedback control module. The method includes:

[0005] A single-frequency laser outputs single-frequency light to an interferometric fiber optic sensor. The data acquisition card applies PGC high-frequency carrier modulation to the single-frequency light or the interferometric fiber optic sensor through internal or external modulation. The interferometric fiber optic sensor outputs interferometric light carrying the physical parameters to be measured.

[0006] The photodetector converts the interference light into an electrical signal, the data acquisition card converts the electrical signal into a normalized long-time domain light intensity digital signal, and transmits it to the signal processing and demodulation device.

[0007] The signal processing and demodulation device divides the normalized long-time-domain light intensity digital signal into several normalized short-time-domain light intensity signals with a length of a single PGC modulation period, which serve as input data for the random forest regression model and the support vector machine regression model optimized by particle swarm optimization. It then performs a Fourier transform on one of the normalized short-time-domain light intensity signals to obtain a frequency domain feature signal, which is then input into the trained random forest regression model to output the PGC modulation depth of the current system.

[0008] The feedback control module adjusts the PGC modulation voltage of the data acquisition card in real time according to the PGC modulation depth, stabilizes the PGC modulation depth of the system to a fixed value, and obtains the normalized short-time domain light intensity signal after parameter stabilization.

[0009] The signal processing and demodulation device inputs the normalized short-time domain light intensity signal after all parameters have been stabilized into the trained particle swarm optimization support vector machine regression model, and outputs the initial phase value corresponding to each normalized short-time domain light intensity signal; it concatenates all the initial phase values ​​in chronological order to obtain the phase signal containing the DC constant term; it performs a DC removal operation on the phase signal containing the DC constant term to eliminate the DC constant term introduced by the system phase delay, and demodulates the physical parameter signal to be measured.

[0010] A machine learning-based PGC light intensity and phase mapping demodulation device, comprising a single-frequency laser, an interferometric fiber optic sensor, a photodetector, a data acquisition card, and a signal processing and demodulation device connected in sequence, wherein the signal processing and demodulation device incorporates a random forest regression model, a particle swarm optimization support vector machine regression model, and a feedback control module, including:

[0011] A single-frequency laser outputs single-frequency light to an interferometric fiber optic sensor. The data acquisition card applies PGC high-frequency carrier modulation to the single-frequency light or the interferometric fiber optic sensor through internal or external modulation. The interferometric fiber optic sensor outputs interferometric light carrying the physical parameters to be measured.

[0012] The photodetector converts the interference light into an electrical signal, the data acquisition card converts the electrical signal into a normalized long-time domain light intensity digital signal, and transmits it to the signal processing and demodulation device.

[0013] The signal processing and demodulation device divides the normalized long-time-domain light intensity digital signal into several normalized short-time-domain light intensity signals with a length of a single PGC modulation period, which serve as input data for the random forest regression model and the support vector machine regression model optimized by particle swarm optimization. It then performs a Fourier transform on any one of the normalized short-time-domain light intensity signals to obtain a frequency domain feature signal, which is then input into the trained random forest regression model to output the PGC modulation depth of the current system.

[0014] The feedback control module adjusts the PGC modulation voltage of the data acquisition card in real time according to the PGC modulation depth, stabilizes the PGC modulation depth of the system to a fixed value, and obtains the normalized short-time domain light intensity signal after parameter stabilization.

[0015] The signal processing and demodulation device inputs the normalized short-time domain light intensity signal after all parameters have been stabilized into the trained particle swarm optimization support vector machine regression model, and outputs the initial phase value corresponding to each normalized short-time domain light intensity signal; it concatenates all the initial phase values ​​in chronological order to obtain the phase signal containing the DC constant term; it performs a DC removal operation on the phase signal containing the DC constant term to eliminate the DC constant term introduced by the system phase delay, and demodulates the physical parameter signal to be measured.

[0016] The aforementioned PGC light intensity and phase mapping demodulation method and apparatus based on machine learning utilizes a random forest regression model to achieve rapid and accurate calculation of modulation depth. This model takes the frequency domain characteristics of the normalized short-time-domain light intensity signal as input and can output accurate modulation depth without relying on phase delay-related parameters. Furthermore, the feedback control module adjusts the modulation voltage in real time based on the result to stabilize the system modulation depth to a fixed value, completely avoiding the problem of repeated and complex calculations of modulation depth as the environment changes in traditional methods. At the same time, this application does not perform any calculation operation on phase delay in the demodulation process. Instead, it allows the particle swarm optimization support vector machine regression model to directly output the initial phase and concatenate it into a phase signal containing a DC constant term. Then, the DC removal operation is used to eliminate the influence of system phase delay. Since different system phase delays only introduce corresponding DC constant terms into the demodulation result, this replaces the complex calculation of phase delay in traditional methods, achieving the effect of ensuring the accuracy of demodulation results without calculating phase delay. Meanwhile, by refactoring the process, traditional computationally intensive demodulation steps such as carrier signal mixing, low-pass filtering, arctangent, or differential multiplication are abandoned. Instead, the normalized long-time-domain digital light intensity signal is segmented into a short-time-domain signal of a single PGC modulation period, which is directly input into a trained particle swarm optimization support vector machine regression model to achieve direct mapping demodulation from light intensity to initial phase. The entire process replaces the complex multi-step signal calculations in traditional methods with the nonlinear mapping of the machine learning model, greatly simplifying the core demodulation process. At the same time, the computational load of operations such as signal segmentation, Fourier transform, model prediction, and DC removal is far lower than the multi-step calculations in traditional demodulation. Especially in large-scale interferometric fiber optic sensor array systems, this simplified process can effectively reduce the computational resources required for demodulating each sensor signal, thereby reducing the power consumption of the entire system and achieving a significant reduction in demodulation overhead from the process execution level. Furthermore, the input and output of each stage in the demodulation process are highly integrated. The prediction of the machine learning model and the adjustment of feedback control form a closed loop. The direct mapping of light intensity and phase and the combination of DC removal operation form an efficient phase extraction method. This allows the entire demodulation process to be free from the complex parameter calculation and signal processing steps of traditional methods, ensuring both the accuracy and robustness of demodulation and fundamentally solving the two core technical problems of traditional PGC demodulation methods. Attached Figure Description

[0017] Figure 1 This is an application diagram of a machine learning-based PGC light intensity and phase mapping demodulation method in one embodiment;

[0018] Figure 2 This is a schematic diagram illustrating the specific application process of the PGC light intensity and phase mapping demodulation method based on machine learning in one embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] In one embodiment, such as Figure 1 As shown, a machine learning-based PGC intensity and phase mapping demodulation method is provided, applied to an interferometric fiber optic sensing system consisting of a single-frequency laser 1, an interferometric fiber optic sensor 2, a photodetector 3, a data acquisition card 4, and a signal processing and demodulation device 5 connected in sequence. The signal processing and demodulation device incorporates a random forest regression model, a particle swarm optimization support vector machine regression model, and a feedback control module, including:

[0021] The single-frequency laser 1 outputs single-frequency light to the interferometric fiber optic sensor 2. The data acquisition card 4 applies PGC high-frequency carrier modulation to the single-frequency light or the interferometric fiber optic sensor through internal modulation or external modulation. The interferometric fiber optic sensor 2 outputs interferometric light carrying the physical parameters to be measured.

[0022] A single-frequency laser 1 provides stable single-frequency input light to the interferometric fiber optic sensor 2. The input light is split into two beams within the sensor, and interference light is generated due to the optical path difference. The PGC internal modulation is achieved by the data acquisition card directly applying high-frequency carrier modulation to the output light of the single-frequency laser. The external modulation is achieved by the data acquisition card applying high-frequency carrier modulation to the piezoelectric ceramic wound around the interferometer arm of the sensor. Phase modulation is achieved by driving the interferometer arm through the piezoelectric ceramic. The external physical parameter to be measured will modulate the interference light, so that the interference light carries the phase information of the parameter to be measured. Finally, the sensor outputs interference light carrying the physical parameter to be measured. The interferometric fiber optic sensor is a Michelson interferometer or a Mach-Zehnder interferometer.

[0023] The photodetector 3 converts the interference light into an electrical signal, and the data acquisition card 4 converts the electrical signal into a normalized long-time domain light intensity digital signal and transmits it to the signal processing and demodulation device.

[0024] The photodetector 3 is the core device for photoelectric conversion. Its core function is to convert the optical signal output from the interferometric fiber optic sensor into a corresponding electrical signal, realizing mode conversion from optical to electrical signal. The data acquisition card 4 performs analog-to-digital conversion on the electrical signal, converting the analog electrical signal into a digital signal, and then normalizes the digital signal to obtain a normalized long-time domain optical intensity digital signal. This signal can be represented as a mathematical expression containing DC term, AC term amplitude, PGC high-frequency carrier modulation depth, PGC high-frequency carrier modulation frequency, system phase delay, and the measured physical parameter signal. The data acquisition card transmits this digital signal to the signal processing and demodulation device as the raw data for subsequent demodulation processing. S ( tIt can be expressed as equation (1):

[0025] (1)

[0026] A express S ( t DC term, B express S ( t The amplitude of the exchange term, C For PGC high-frequency carrier modulation depth, ω 0 represents the PGC high-frequency carrier modulation frequency. θ c For system phase delay, The signal is the physical parameter to be measured.

[0027] In traditional PGC demodulation methods, S ( t High-frequency signal mixing, low-pass filtering, arctangent calculation, or differential multiplication are required to demodulate the measured physical parameter signal. The complex demodulation process in traditional demodulation methods results in large computational resources and high power consumption in large-scale interferometric fiber optic sensor array systems.

[0028] The signal processing and demodulation device 5 divides the normalized long-time-domain light intensity digital signal into several normalized short-time-domain light intensity signals with a length of a single PGC modulation period, which are used as input data for the random forest regression model and the support vector machine regression model optimized by particle swarm optimization. It then performs a Fourier transform on one of the normalized short-time-domain light intensity signals to obtain a frequency domain feature signal, inputs the frequency domain feature signal into the trained random forest regression model, and outputs the PGC modulation depth of the current system.

[0029] like Figure 2 As shown, the signal processing and demodulation device 5 first performs data segmentation processing on the normalized long-time-domain optical intensity digital signal, dividing it into several normalized short-time-domain optical intensity signals with a length of one PGC modulation period. The length of the short-time-domain signal is an integer multiple of the PGC modulation period, typically one PGC modulation period. Then, any one of the normalized short-time-domain optical intensity signals is selected to perform a Fourier transform, converting the time-domain signal into a frequency-domain feature signal. This frequency-domain feature signal contains key spectral features such as the main frequency amplitude, harmonic component ratio, and spectral energy distribution. The random forest regression model is a trained machine learning model that integrates several decision trees. It can receive the frequency-domain feature signal as input and complete the nonlinear mapping from frequency-domain features to PGC modulation depth through the collaborative operation of multiple decision trees. Finally, it outputs the PGC modulation depth of the current system. Moreover, the trained model has generalization ability for sensing systems with different system phase delays, and changes in system phase delay will not affect the calculation result of PGC modulation depth.

[0030] The feedback control module adjusts the PGC modulation voltage of the data acquisition card in real time according to the PGC modulation depth, stabilizes the PGC modulation depth of the system to a fixed value, and obtains the normalized short-time domain light intensity signal after parameter stabilization.

[0031] The feedback control module is a closed-loop control module built into the signal processing demodulation device. Its input is connected to the output of the random forest regression model, and it can receive the system PGC modulation depth calculated by the model in real time. Based on the deviation between the real-time value and the fixed value of the modulation depth, this module dynamically adjusts the PGC modulation voltage of the data acquisition card to achieve closed-loop calibration of the system PGC modulation depth, so that the modulation depth is stabilized at a fixed value (typically 2.63 rad), thereby obtaining the normalized short-time domain light intensity signal after parameter stabilization. The core purpose of this operation is to reduce the training cost of the support vector machine regression model optimized by particle swarm optimization, avoid the sensitivity of PGC demodulation results to demodulation parameters, and enhance the robustness of the demodulation process of the interferometric fiber optic sensing system.

[0032] The signal processing and demodulation device 5 inputs the normalized short-time domain light intensity signal after all parameters have been stabilized into the trained particle swarm optimization support vector machine regression model, and outputs the initial phase value corresponding to each normalized short-time domain light intensity signal; it concatenates all the initial phase values ​​in time order to obtain the phase signal containing the DC constant term; it performs a DC removal operation on the phase signal containing the DC constant term to eliminate the DC constant term introduced by the system phase delay, and demodulates the physical parameter signal to be measured.

[0033] like Figure 2 As shown, the particle swarm optimization support vector machine regression model is a trained light intensity-phase mapping core model. It can receive normalized short-time domain light intensity signals with stable parameters as input and directly output the initial phase value corresponding to each signal. After concatenating all the initial phase values ​​in time order, the resulting phase signal contains a DC constant term introduced by the system phase delay. This DC constant term is a single fixed value, determined only by the system phase delay and independent of the physical parameter to be measured. After eliminating this DC constant term through a de-DC operation, the signal of the physical parameter to be measured can be demodulated.

[0034] Specifically, the trained model can directly and independently output the initial phase corresponding to each segment of light intensity signal based on the waveform characteristics of each short-time domain signal. The signal can be obtained by sequentially concatenating the initial phases corresponding to consecutive short-time domain light intensity signals. , The expression is as shown in equation (2). The DC constant term introduced for the phase delay is used to demodulate the measured physical parameter signal through a DC removal operation. ;

[0035] (2)

[0036] Traditional PGC demodulation methods require complex calculations to determine the system modulation depth and phase delay to avoid errors in the demodulation results. However, in this application, different system phase delays only introduce a corresponding DC constant term into the demodulation results. The influence of the phase delay parameter on the demodulation results can be eliminated through a DC removal operation. This process does not require any complex calculations of the system phase delay, and the measured physical parameter can be one or more of sound, vibration, or temperature.

[0037] Specifically, such as Figure 2 As shown, in the PGC internal modulation scheme, the data acquisition card 4 performs high-frequency carrier phase modulation on the output light of the single-frequency laser 1. Alternatively, in the PGC external modulation scheme, one interferometer arm of the interferometric fiber optic sensor 2 is wound with piezoelectric ceramic, and the data acquisition card 4 performs high-frequency carrier phase modulation on the interferometer arm through the piezoelectric ceramic. The output end of the single-frequency laser 1 outputs single-frequency light to the input end of the interferometric fiber optic sensor 2. The interferometric fiber optic sensor 2 outputs output light modulated by the high-frequency carrier phase and the external environmental signal to be measured. The output end of the interferometric fiber optic sensor 2 is connected to the input end of the photodetector 3. The photodetector 3 converts the optical signal output by the interferometric fiber optic sensor 2 into an electrical signal. The input end of the data acquisition card 4 is connected to the output end of the photodetector 3. The data acquisition card 4 converts the electrical signal output by the photodetector 3 into a digital signal and transmits it to the signal processing and demodulation device 5. This digital signal is the time-domain output light intensity signal of the interferometric fiber optic sensor 2. The signal processing and demodulation device 5 normalizes the long-time-domain light intensity signal output from the interferometric fiber optic sensor 2 and then segments the data into several short-time-domain signals with single PGC cycles. Each short-time-domain signal is Fourier transformed into a frequency-domain signal and then fed into a trained random forest regression model. This model outputs the system's PGC modulation depth in real time and accurately. The feedback control module controls the modulation voltage of the data acquisition card based on the PGC modulation depth calculated by the random forest regression model, thereby stabilizing the system's PGC modulation depth at 2.63 rad (or other fixed modulation depth). Simultaneously, the signal processing and demodulation device 5 feeds several normalized, continuous short-time-domain light intensity signals with single PGC cycles into a particle swarm optimization support vector machine regression model. The model directly and independently outputs the initial phase calculation value corresponding to each light intensity signal segment. After sequentially splicing the calculated initial phases and removing DC, the measured signal can be demodulated.

[0038] In one embodiment, internal modulation is achieved by the data acquisition card directly applying PGC high-frequency carrier modulation to the single-frequency light output by the single-frequency laser; external modulation is achieved by the data acquisition card applying PGC high-frequency carrier modulation to the piezoelectric ceramic wound on the interferometric arm of the interferometric fiber sensor, thereby driving the interferometric arm through the piezoelectric ceramic to achieve phase modulation.

[0039] Specifically, the PGC modulation scheme of this application includes two implementation forms: internal modulation and external modulation. Both forms are driven by a data acquisition card providing the driving signal for PGC high-frequency carrier modulation, adapting to different application scenarios of interferometric fiber optic sensing systems. The internal modulation scheme eliminates the hardware deployment of piezoelectric ceramics and directly modulates the phase of the output light of the single-frequency laser, resulting in a shorter modulation link and less signal loss. The external modulation scheme achieves phase modulation by driving the interferometer arm with piezoelectric ceramics to change the optical path difference, resulting in more accurate modulation and adapting to sensing scenarios with high modulation accuracy requirements. The core purpose of both modulation schemes is to apply PGC high-frequency carrier modulation to the interferometric light, laying the foundation for subsequent intensity-phase mapping demodulation. Both schemes can be seamlessly adapted to subsequent machine learning demodulation models and feedback control modules without affecting demodulation performance and robustness.

[0040] In one embodiment, the physical parameter to be measured is one or more of sound, vibration, or temperature.

[0041] Specifically, the core sensing principle of the interferometric fiber optic sensor is that changes in external physical parameters alter the optical path difference of the interferometric light within the sensor, thereby causing a change in the phase of the interferometric light. This application utilizes this sensing characteristic to detect various environmental physical parameters such as sound, vibration, and temperature. When detecting sound parameters, the vibration of the sound wave drives the sensor's interferometer arm to undergo slight deformation, changing the optical path difference. When detecting vibration parameters, external vibration directly acts on the sensor, causing a corresponding change in the phase of the interferometric light. When detecting temperature parameters, temperature changes cause the optical fiber to expand and contract, changing the optical path difference and achieving phase modulation. The demodulation method of this application can accurately extract the phase changes caused by the aforementioned physical parameters through intensity-phase mapping, without requiring adjustments to the demodulation model and parameters for different parameters. It has good versatility and can be widely applied in various fiber optic sensing scenarios such as industrial monitoring, environmental sensing, and security detection.

[0042] In one embodiment, the training process of the random forest regression model is as follows: normalized short-time domain light intensity signals output by interferometric fiber optic sensors at different PGC modulation depths are collected, and frequency domain feature signals are obtained by Fourier transform. A sample dataset of frequency domain feature signals and PGC modulation depths is constructed. The sample dataset is input into the random forest regression model for training, so that the model learns the nonlinear mapping relationship between the frequency domain feature signals and the PGC modulation depth. The trained model has generalization ability for sensing systems with different system phase delays, and the system phase delay does not affect the calculation result of the PGC modulation depth.

[0043] Specifically, the training of the random forest regression model uses PGC modulation depth as the prediction target and frequency domain feature signals as input features. Before training, multiple sets of normalized short-time domain light intensity signals under different PGC modulation depths need to be collected. Fourier transforms are performed on each signal to obtain the corresponding frequency domain feature signals, constructing a one-to-one frequency domain feature signal-PGC modulation depth sample dataset. During training, several decision trees within the model learn the intrinsic correlation between spectral features and modulation depth based on different feature subsets and data subsets, using a threshold-based tree splitting rule. The model outputs a preliminary prediction of the modulation depth at the leaf nodes. Finally, it aggregates and averages the outputs of all decision trees to obtain the accurate prediction of the PGC modulation depth. Since the input of the model is a frequency domain feature signal rather than a time domain signal, the system phase delay only affects the phase characteristics of the time domain signal and does not change the spectral characteristics of the frequency domain feature signal. Therefore, the trained model has good generalization ability for sensing systems with different system phase delays. There is no need to retrain the model for different phase delays, which greatly improves the model's adaptability. At the same time, it saves the calculation steps of the system phase delay and reduces the demodulation complexity.

[0044] In one embodiment, the training process of the particle swarm optimization support vector machine regression model is as follows: at a fixed PGC modulation depth, the normalized short-time domain light intensity signal output by the interferometric fiber optic sensor and its corresponding true initial phase value are collected to construct a sample dataset of light intensity signal-initial phase; the particle swarm optimization algorithm is used to globally optimize the kernel function parameters and penalty factor of the support vector machine regression model to determine the optimal hyperparameter combination; the sample dataset is input into the support vector machine regression model with the optimal hyperparameter combination for training, so that the model learns the nonlinear mapping relationship between the normalized short-time domain light intensity signal and the initial phase.

[0045] Specifically, the training of the particle swarm optimization support vector machine regression model needs to be completed at a fixed PGC modulation depth (typically 2.63 rad), a prerequisite guaranteed by the feedback control module during the actual demodulation process. Before training, normalized short-time domain light intensity signals at this fixed modulation depth need to be collected, and the true initial phase value corresponding to each signal needs to be obtained to construct a sample dataset of light intensity signal-initial phase. Since the demodulation effect of the support vector machine regression model is significantly affected by the kernel function parameters and penalty factors, this application uses a particle swarm optimization algorithm to globally optimize these hyperparameters, finding the optimal hyperparameters through iterative search of the particle swarm to minimize model prediction errors. The optimal hyperparameter combination with the smallest difference solves the problem of blind selection in traditional hyperparameters. After the sample dataset is input into the support vector machine regression model with the optimal hyperparameter combination, the model projects the time-domain light intensity signal onto a high-dimensional feature space through nonlinear mapping. In this space, the optimal regression hyperplane that minimizes the deviation between the predicted initial phase and the true initial phase is constructed, thereby learning the complex nonlinear mapping relationship between the normalized short-time-domain light intensity signal and the initial phase. The trained model can directly output the corresponding initial phase value independently based on the waveform morphology characteristics of the light intensity signal, without additional signal processing steps, greatly simplifying the demodulation process.

[0046] In one embodiment, the DC removal operation employs the mean-subtraction method, which calculates the average value of the spliced ​​phase signal containing the DC constant term, and subtracts the average value from each phase value to obtain the measured physical parameter signal with the DC constant term removed.

[0047] Specifically, the system phase delay introduces only a single DC constant term into the spliced ​​phase signal. This constant term is a fixed value throughout the entire phase signal and does not change with time or the measured physical parameter. Therefore, efficient DC removal can be achieved through the mean-subtraction method. In practice, the arithmetic mean of all phase values ​​in the phase signal containing the DC constant term is first calculated. This average value is the equivalent value of the DC constant term introduced by the system phase delay. Then, each phase value is subtracted from this average value, and the difference obtained is the phase signal with the DC constant term removed. This signal is the final demodulated measured physical parameter signal. This DC removal method has low computational complexity and simple operation, and can be quickly implemented in signal processing demodulation devices. It can completely eliminate the influence of system phase delay on the demodulation result. Compared with the complex phase delay calculation method in traditional PGC demodulation, it significantly reduces computational resource consumption, improves the real-time performance of demodulation, and also eliminates the need to consider the differences in system phase delay, thus improving generalization.

[0048] In one embodiment, the particle swarm optimization support vector machine regression model treats the normalized short-time domain light intensity signal of a single PGC modulation period as an independent phase prediction unit, with each unit corresponding to an initial phase value.

[0049] Specifically, this application uses a single PGC modulation period as the basic unit for phase prediction. A particle swarm optimization support vector machine regression model performs independent phase prediction for each normalized short-time-domain light intensity signal of this length, with one initial phase value output for each short-time-domain signal. This design reduces the equivalent signal sampling rate and probe bandwidth of the demodulation result to 1 / 3 of that of traditional PGC demodulation methods. fc ( fc Using the PGC modulation frequency, while ensuring effective demodulation of the measured physical parameters, it significantly reduces the amount of data processed by the signal, and reduces the computational load and power consumption of the signal processing and demodulation device. It is especially suitable for large-scale interferometric fiber optic sensor array systems, and can effectively solve the technical problems of limited computing resources and excessive power consumption in array demodulation. At the same time, using a single PGC modulation period as the prediction unit ensures that each predicted value contains complete PGC modulation features, avoids feature loss caused by signal segmentation, and ensures the accuracy of the demodulation results.

[0050] In one embodiment, the normalized long-time-domain optical intensity digital signal is segmented in a non-overlapping equal-length manner. The time length of each normalized short-time-domain optical intensity signal after segmentation is completely consistent with the period of PGC high-frequency carrier modulation, ensuring that each short-time-domain signal contains complete single-cycle PGC modulation features.

[0051] Specifically, to match the training and prediction rules of the support vector machine regression model optimized by particle swarm optimization, the signal processing demodulation device processes the normalized long-time-domain optical intensity digital signal using a non-overlapping, equal-length segmentation method. The segmentation time length is completely consistent with the period of PGC high-frequency carrier modulation, ensuring that each normalized short-time-domain optical intensity signal after segmentation contains complete single-cycle PGC modulation features, avoiding demodulation errors caused by feature loss. The non-overlapping segmentation method avoids redundant data processing, reduces computation, and improves demodulation efficiency. It also ensures that each short-time-domain signal is continuous in time and without redundancy, guaranteeing that the spliced ​​phase signal can truly reflect the changing pattern of the measured physical parameters. This segmentation method can be quickly implemented through software programming or hardware logic circuits of the signal processing demodulation device, adapting to different signal processing equipment. Furthermore, the segmented signal can be directly used as input to two machine learning models without additional preprocessing steps, simplifying the demodulation process.

[0052] In one embodiment, the signal processing demodulation device is a data processing computer or an FPGA development board. The random forest regression model and the particle swarm optimization support vector machine regression model are deployed in the signal processing demodulation device through software programming or hardware logic circuits to realize real-time parallel processing of PGC modulation depth calculation and light intensity-phase mapping.

[0053] Specifically, the signal processing and demodulation device is the core execution carrier of the demodulation method of this application. It can be a data processing computer or an FPGA development board, adapting to different application scenarios: the data processing computer is suitable for demodulation scenarios of small-scale sensing systems with high deployment flexibility requirements, and can deploy two machine learning models through software programming such as Python and Matlab; the FPGA development board is suitable for large-scale interferometric fiber optic sensor array systems with high requirements for real-time performance, power consumption, and integration, and can solidify the model through hardware logic circuits to achieve fast hardware-level demodulation; the two machine learning models can be processed in real time in parallel in the signal processing and demodulation device, that is, while the random forest regression model calculates the PGC modulation depth, the particle swarm optimization support vector machine regression model can simultaneously perform intensity-phase mapping demodulation without serial waiting, which greatly improves the real-time performance of demodulation and meets the detection requirements of fiber optic sensing systems for dynamic physical parameters to be measured; at the same time, the software or hardware deployment of the model has good portability and scalability, and can be flexibly adjusted according to the needs of the actual sensing system, which improves the practicality of the demodulation method of this application.

[0054] In one embodiment, a machine learning-based PGC light intensity and phase mapping demodulation device is provided, comprising: a single-frequency laser, an interferometric fiber optic sensor, a photodetector, a data acquisition card, and a signal processing demodulation device connected in sequence, wherein the signal processing demodulation device incorporates a random forest regression model, a particle swarm optimization support vector machine regression model, and a feedback control module.

[0055] A single-frequency laser outputs single-frequency light to an interferometric fiber optic sensor. The data acquisition card applies PGC high-frequency carrier modulation to the single-frequency light or the interferometric fiber optic sensor through internal or external modulation. The interferometric fiber optic sensor outputs interferometric light carrying the physical parameters to be measured.

[0056] The photodetector converts the interference light into an electrical signal, the data acquisition card converts the electrical signal into a normalized long-time domain light intensity digital signal, and transmits it to the signal processing and demodulation device.

[0057] The signal processing and demodulation device divides the normalized long-time-domain light intensity digital signal into several normalized short-time-domain light intensity signals with a length of a single PGC modulation period, which serve as input data for the random forest regression model and the support vector machine regression model optimized by particle swarm optimization. It then performs a Fourier transform on one of the normalized short-time-domain light intensity signals to obtain a frequency domain feature signal, which is then input into the trained random forest regression model to output the PGC modulation depth of the current system.

[0058] The feedback control module adjusts the PGC modulation voltage of the data acquisition card in real time according to the PGC modulation depth, stabilizes the PGC modulation depth of the system to a fixed value, and obtains the normalized short-time domain light intensity signal after parameter stabilization.

[0059] The signal processing and demodulation device inputs the normalized short-time domain light intensity signal after all parameters have been stabilized into the trained particle swarm optimization support vector machine regression model, and outputs the initial phase value corresponding to each normalized short-time domain light intensity signal; it concatenates all the initial phase values ​​in chronological order to obtain the phase signal containing the DC constant term; it performs a DC removal operation on the phase signal containing the DC constant term to eliminate the DC constant term introduced by the system phase delay, and demodulates the physical parameter signal to be measured.

[0060] Specific limitations regarding the machine learning-based PGC intensity and phase mapping demodulation device can be found in the limitations of the machine learning-based PGC intensity and phase mapping demodulation method described above, and will not be repeated here. Each module in the aforementioned machine learning-based PGC intensity and phase mapping demodulation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0061] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0062] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0063] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A machine learning-based PGC light intensity and phase mapping demodulation method, characterized in that, An interferometric fiber optic sensing system is applied to a system consisting of a single-frequency laser, an interferometric fiber optic sensor, a photodetector, a data acquisition card, and a signal processing and demodulation device connected in sequence. The signal processing and demodulation device incorporates a random forest regression model, a particle swarm optimization support vector machine regression model, and a feedback control module. The method includes: A single-frequency laser outputs single-frequency light to an interferometric fiber optic sensor. The data acquisition card applies PGC high-frequency carrier modulation to the single-frequency light or the interferometric fiber optic sensor through internal or external modulation. The interferometric fiber optic sensor outputs interferometric light carrying the physical parameters to be measured. The photodetector converts the interference light into an electrical signal, the data acquisition card converts the electrical signal into a normalized long-time domain light intensity digital signal, and transmits it to the signal processing and demodulation device. The signal processing and demodulation device divides the normalized long-time-domain light intensity digital signal into several normalized short-time-domain light intensity signals with a length of a single PGC modulation period, which are used as input data for the random forest regression model and the particle swarm optimization support vector machine regression model. The device then performs a Fourier transform on one of the normalized short-time-domain light intensity signals to obtain a frequency domain feature signal, which is then input into the trained random forest regression model to output the PGC modulation depth of the current system. The feedback control module adjusts the PGC modulation voltage of the data acquisition card in real time according to the PGC modulation depth, stabilizes the PGC modulation depth of the system to a fixed value, and obtains the normalized short-time domain light intensity signal after parameter stabilization. The signal processing and demodulation device inputs the normalized short-time domain light intensity signal after all parameters have been stabilized into the trained particle swarm optimization support vector machine regression model, and outputs the initial phase value corresponding to each normalized short-time domain light intensity signal; it concatenates all the initial phase values ​​in chronological order to obtain a phase signal containing a DC constant term; it performs a DC removal operation on the phase signal containing the DC constant term to eliminate the DC constant term introduced by the system phase delay, and demodulates the physical parameter signal to be measured.

2. The method according to claim 1, characterized in that, The internal modulation is achieved by the data acquisition card directly applying PGC high-frequency carrier modulation to the single-frequency light output by the single-frequency laser; the external modulation is achieved by the data acquisition card applying PGC high-frequency carrier modulation to the piezoelectric ceramic wound on the interferometer arm of the interferometric fiber sensor, and phase modulation is achieved by driving the interferometer arm through the piezoelectric ceramic.

3. The method according to claim 1, characterized in that, The physical parameter to be measured is one or more of sound, vibration, or temperature.

4. The method according to claim 1, characterized in that, The training process of the random forest regression model is as follows: normalized short-time domain light intensity signals output by interferometric fiber optic sensors at different PGC modulation depths are collected, and frequency domain feature signals are obtained by Fourier transform. A sample dataset of frequency domain feature signals and PGC modulation depths is constructed. The sample dataset is input into the random forest regression model for training, so that the model learns the nonlinear mapping relationship between frequency domain feature signals and PGC modulation depths.

5. The method according to claim 1, characterized in that, The training process of the particle swarm optimization support vector machine regression model is as follows: Under a fixed PGC modulation depth, the normalized short-time domain light intensity signal output by the interferometric fiber optic sensor and its corresponding true initial phase value are collected to construct a sample dataset of light intensity signal-initial phase; the particle swarm optimization algorithm is used to globally optimize the kernel function parameters and penalty factor of the support vector machine regression model to determine the optimal hyperparameter combination; the sample dataset is input into the support vector machine regression model with the optimal hyperparameter combination for training, so that the model learns the nonlinear mapping relationship between the normalized short-time domain light intensity signal and the initial phase.

6. The method according to claim 1, characterized in that, The DC removal operation employs the mean-subtraction method, which calculates the average value of the spliced ​​phase signal containing the DC constant term, and subtracts the average value from each phase value to obtain the measured physical parameter signal with the DC constant term removed.

7. The method according to claim 1, characterized in that, The particle swarm optimization support vector machine regression model treats the normalized short-time domain light intensity signal of a single PGC modulation period as an independent phase prediction unit, with each unit corresponding to an initial phase value.

8. The method according to claim 1, characterized in that, The normalized long-time-domain optical intensity digital signal is segmented in a non-overlapping, equal-length manner. The time length of each normalized short-time-domain optical intensity signal after segmentation is completely consistent with the period of PGC high-frequency carrier modulation, ensuring that each short-time-domain signal contains complete single-cycle PGC modulation features.

9. The method according to claim 1, characterized in that, The signal processing and demodulation device is a data processing computer or an FPGA development board. The random forest regression model and the particle swarm optimization support vector machine regression model are deployed in the signal processing and demodulation device through software programming or hardware logic circuits to realize real-time parallel processing of PGC modulation depth calculation and light intensity-phase mapping.

10. A machine learning-based PGC light intensity and phase mapping demodulation device, characterized in that, The device includes a single-frequency laser, an interferometric fiber optic sensor, a photodetector, a data acquisition card, and a signal processing and demodulation device connected in sequence. The signal processing and demodulation device has a built-in random forest regression model, a particle swarm optimization support vector machine regression model, and a feedback control module. A single-frequency laser outputs single-frequency light to an interferometric fiber optic sensor. The data acquisition card applies PGC high-frequency carrier modulation to the single-frequency light or the interferometric fiber optic sensor through internal or external modulation. The interferometric fiber optic sensor outputs interferometric light carrying the physical parameters to be measured. The photodetector converts the interference light into an electrical signal, the data acquisition card converts the electrical signal into a normalized long-time domain light intensity digital signal, and transmits it to the signal processing and demodulation device. The signal processing and demodulation device divides the normalized long-time-domain light intensity digital signal into several normalized short-time-domain light intensity signals with a length of a single PGC modulation period, which are used as input data for the random forest regression model and the particle swarm optimization support vector machine regression model. The device then performs a Fourier transform on one of the normalized short-time-domain light intensity signals to obtain a frequency domain feature signal, which is then input into the trained random forest regression model to output the PGC modulation depth of the current system. The feedback control module adjusts the PGC modulation voltage of the data acquisition card in real time according to the PGC modulation depth, stabilizes the PGC modulation depth of the system to a fixed value, and obtains the normalized short-time domain light intensity signal after parameter stabilization. The signal processing and demodulation device inputs the normalized short-time domain light intensity signal after all parameters have been stabilized into the trained particle swarm optimization support vector machine regression model, and outputs the initial phase value corresponding to each normalized short-time domain light intensity signal; it concatenates all the initial phase values ​​in chronological order to obtain a phase signal containing a DC constant term; it performs a DC removal operation on the phase signal containing the DC constant term to eliminate the DC constant term introduced by the system phase delay, and demodulates the physical parameter signal to be measured.

Images