Machine learning assisted fiber optic gyroscope polarization locking and slow drift compensation system and method

Abstract

The application discloses a machine learning assisted fiber-optic gyroscope polarization locking and slow drift compensation system. The system can intelligently search and predict the relationship between the fiber-optic gyroscope interference error signal and the incident light polarization direction by introducing the neural network in machine learning, and feed back the predicted optimal value to the hardware system for compensation, thereby realizing the functions of fiber polarization locking and suppressing the intensity slow drift of the output signal. The system comprises a coherent light source, a Glan laser prism, a first polarization beam splitter, a first mirror, a first Faraday rotator, a first half-wave plate, a liquid crystal modulator, a polarization maintaining optical fiber, a phase modulator, a second Faraday rotator, a second half-wave plate, a second mirror, a third half-wave plate, a second polarization beam splitter, a photodetector, a computer and a signal generator. The application also discloses a machine learning assisted fiber-optic gyroscope polarization locking and slow drift compensation method realized by using the above system.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of artificial intelligence, electronic control, quantum optics and precision measurement, and mainly relates to a machine learning assisted polarization locking and slow drift compensation system and method of an optical fiber gyroscope, which is used for angular velocity measurement and phase measurement of the optical fiber gyroscope. BACKGROUND

[0002] The optical fiber gyroscope has important significance in the field of precision measurement and has broad application prospects in the fields of gravitational wave detection, distributed sensing, angular velocity measurement, space navigation, temperature sensing, deformation detection, etc. It can overcome some shortcomings of traditional mechanical gyroscopes, suspended gyroscopes and piezoelectric gyroscopes by using laser means. In addition, thanks to the maturity of optical fiber micro-nano processing technology, multiple turns of polarization maintaining optical fiber coils can be used to significantly increase the loop length, further improving the angular velocity measurement accuracy. However, due to the introduction of optical fiber, the gyroscope inevitably brings errors such as optical fiber slow drift, zero drift and noise. Especially for interferometric fiber-optic gyroscopes, slow drift and noise can come from temperature, mechanical disturbance, stress deformation, polarization coupling, Brillouin scattering, Rayleigh scattering, Kerr effect and many other factors. Among them, temperature, mechanical disturbance and polarization mode coupling are important sources of noise and also the main source of slow drift of the optical fiber gyroscope signal. And under the conditions of relatively constant control of experimental temperature and platform isolation, the polarization factor becomes the object that needs to be considered.

[0003] The measurement accuracy of angular velocity and the robustness of the gyroscope are the key to the accuracy of space navigation and positioning, therefore, people need to eliminate unnecessary noise and slow drift as much as possible to improve the angular velocity measurement accuracy and robustness of the optical fiber gyroscope. The traditional method can be mainly divided into two categories: one is to start from the optical path structure, such as using minimum reciprocity optical path, integrated Y waveguide or four-pole symmetric ring structure to reduce slow drift and noise, and then improve the accuracy of the gyroscope; the other method is to start from the control algorithm, such as using proportional integral differential algorithm to close loop feedback to phase lock, polarization lock or using Kalman filter method to post-process the collected data. The traditional proportional integral differential algorithm relies on multiple pre-experiments to explore the experience parameters of the controller, and the selection of parameters has a great influence on the locking effect. The data post-processing method cannot directly act on the experimental measurement system itself and cannot be processed quickly and in real time, which greatly reduces the applicability and operability of this method in actual complex and variable environments. SUMMARY

[0004] To address the shortcomings of existing technologies, this invention aims to provide a novel machine learning-assisted polarization locking and drift compensation system and method for fiber optic gyroscopes, specifically designed to solve the problems of polarization mode coupling, slow output signal drift, and low measurement sensitivity in traditional fiber optic gyroscopes. Inspired by traditional interferometer phase-locked loop (PLL) technology, this invention combines machine learning and proportional-integral-derivative (PID) control algorithms, adapting them to the fiber optic gyroscope hardware system. It can intelligently search for the global minimum of the fiber optic gyroscope error signal while providing real-time, rapid response to avoid additional interference with the measured angular velocity signal. The system effectively addresses the impact of polarization factors on the fiber optic gyroscope signal through cross-control of both program and hardware. Compared to traditional fiber optic gyroscopes, this invention significantly suppresses signal drift, greatly improves long-term robustness and measurement sensitivity, and enhances anti-interference capabilities in general environments. This invention further enhances the application value of fiber optic gyroscopes in precision measurement fields such as fiber optic sensing, space navigation, and geographic surveying.

[0005] To achieve the above objectives, this invention designs and implements a machine learning-assisted fiber optic gyroscope polarization locking and slow drift compensation system, as detailed below:

[0006] A coherent light source is a linearly polarized light generated by a continuous laser. After entering the fiber optic loop of a gyroscope, it can sense the phase difference between the clockwise and counterclockwise loops caused by the Sagnac effect.

[0007] Glan laser prisms, with their high extinction ratio, are used to further improve the linear polarization purity of incident coherent light sources, ensuring that the light incident on the loop is linearly polarized with an extinction ratio greater than 1:10000 and that the polarization direction is adjustable.

[0008] The first polarization beam splitter is used to split linearly polarized light into beams 50:50 along the horizontal and vertical polarization directions;

[0009] The first reflecting mirror is used to change the propagation direction of the light field so that the light path coincides with the center line of the subsequent device;

[0010] The first Faraday rotator is a non-reciprocal device that uses the Faraday rotation effect to rotate the polarization direction of light passing in the forward direction by 45 degrees and rotate the polarization direction of light passing in the reverse direction by 45 degrees along the rotation direction when passing in the forward direction, so that the polarization of the forward incident light and the reverse outgoing light are orthogonal to each other.

[0011] The first half-wave plate can be used to adjust the polarization direction of linearly polarized light, specifically to align the polarization direction of the light field with the crystal optical axis of the subsequent liquid crystal modulator.

[0012] Liquid crystal modulator is a core device for hardware compensation feedback, which is mainly composed of liquid crystal retardation wave plate and 1 / 4 wave plate. The slow axis of the two is at an angle of 45 degrees. In the liquid crystal modulator, the liquid crystal molecules can adjust the arrangement under the drive of external voltage, causing the change of refractive index, and then causing the rotation of the polarization direction of the outgoing light. Under the influence of environmental temperature change, vibration and other factors, polarization maintaining optical fiber will deform, optical axis twist, polarization coupling and other phenomena. The role of the liquid crystal modulator is to real-time correct the polarization direction of linearly polarized light under the control of feedback voltage, so that it can be aligned with the optical axis of the polarization maintaining optical fiber for a long time. The specific structure of the liquid crystal modulator is shown in the accompanying Figure 3 When used, the incident light polarization is aligned with the calibration axis on the shell of the liquid crystal modulator, and is incident on the liquid crystal. Voltage is applied along the fast axis or slow axis of the liquid crystal to cause the change of refractive index between the fast and slow axes. After passing through the 1 / 4 wave plate, the polarization angle will rotate.

[0013] The laser sequentially passes through the first half wave plate and the liquid crystal modulator, and is incident into the polarization maintaining optical fiber.

[0014] The polarization maintaining optical fiber constitutes a sagnac loop, so that the sagnac effect is generated when the linearly polarized incident light propagates in the optical fiber medium, and the polarization maintaining optical fiber can ensure that the linearly polarized light incident along the optical axis is as stable as possible, so that the polarization of the outgoing light remains as much as possible;

[0015] The phase modulator can change the refractive index difference between the fast axis and the slow axis under the drive of external voltage, and then generate a phase difference, which is used for phase shifting of interference signals in the fiber optic gyroscope system;

[0016] The second Faraday rotator has a similar effect to the first Faraday rotator, and is used to make the polarization direction of the forward incident light and the reverse outgoing light perpendicular to each other;

[0017] The second half wave plate is used to adjust the polarization direction of the linearly polarized light, and preliminarily aligns the optical axis of the polarization maintaining optical fiber;

[0018] Between the liquid crystal modulator and the polarization maintaining optical fiber, and between the second half wave plate and the polarization maintaining optical fiber, there are head and tail coupling interfaces of the polarization maintaining optical fiber respectively. The coupling interface plays a role of converging the incident light, so that the energy of the laser is as much as possible to be used and transmitted along the optical fiber.

[0019] The second mirror can change the propagation direction of the light, so that the light is collimated and incident on the subsequent device;

[0020] The third half wave plate is used to change the polarization state of the outgoing light from the fiber loop, so that the polarization directions of the clockwise outgoing light and the counterclockwise outgoing light coincide, and the interference condition is met;

[0021] The second polarization beam splitter cooperates with the third half-wave plate in front, so that the light intensity transmitted from the second polarization beam splitter reaches the maximum, which is equivalent to a polarizer;

[0022] The photodetector contains a photodiode for converting optical signals into electrical signals through the photoelectric effect;

[0023] The computer is used to run the acquisition program of the optical signal and the optimization program of machine learning, and control the transmission of data, which plays a role in communication with the hardware system;

[0024] The signal generator is an arbitrary waveform pulse generator, which can be controlled by the computer to combine the input pulse digital signal into an analog electrical signal and transmit it to the liquid crystal modulator through the BNC cable.

[0025] Notes:

[0026] The coherent light source is generated by a continuous laser, with a line width of 1.5 kHz, and the optical power is very stable, and it is linearly polarized light.

[0027] The aforementioned first polarization beam splitter splits the incident light according to 50:50, and the polarization of the two beams of light after splitting is perpendicular to each other.

[0028] The aforementioned liquid crystal modulator receives the electrical signal from the aforementioned signal generator and responds to the optical path, so that the polarization of the optical field is quickly and real-time corrected, and the slow drift of the output signal caused by the polarization factor is suppressed. And according to the response limit of the liquid crystal modulator, the amplitude of the electrical signal it receives must be set within the limited range. The working voltage of the liquid crystal modulator has a limited range of about 0-5V. Within the range of 0-5V, the response of the liquid crystal modulator is relatively sensitive. After exceeding the limited range, the modulation response of the liquid crystal is very slow and there is a risk of voltage breakdown, which will affect the service life. In order to quickly respond to feedback, the amplitude value of the working voltage of the liquid crystal modulator in the experiment must be set within the limited range.

[0029] The clockwise incident light and the counterclockwise incident light can perceive the phase difference caused by the rotation of the external angular velocity when passing through the aforementioned polarization maintaining optical fiber, and the fiber optic gyroscope angular velocity data is obtained according to the formula of the sagnac effect.

[0030] The above-mentioned phase modulator can change the relative phase difference between the clockwise emitted light and the counterclockwise emitted light in the fiber optic gyroscope under the power supply of the high-voltage power supply, and fix the working point of the fiber optic gyroscope near the phase point, that is, near the balance position of the interference signal.

[0031] The aforementioned computer acquires the signal detected by the photoelectric detector and transmits it into the optimization program of machine learning, and outputs it to the aforementioned signal generator after operation via the proportional integral differential operation module. The machine learning program and the proportional integral differential module are mixed and circulated with each other, and the difference between the output signal of the gyroscope interference and the set value is used as the optimization criterion. The above-mentioned feedback system can be stably operated for a long time.

[0032] As shown in Figure 2 , the part in the block on the right side is the specific circulation process of the hybrid algorithm. When the target value collected by the experimental system (the difference between the output signal of the gyroscope interference and the set value) is greater than the set threshold value, it indicates that the gyroscope has deviated from the global optimal value of the working operation at this time, and the neural network algorithm part (machine learning program) in the hybrid algorithm starts to search quickly in the given range to make the target value approach the optimal working point (global optimal value) again. When the optimal target value is found (the end condition of the neural network algorithm is met), at this time, the proportional integral differential PID part in the hybrid algorithm will work, and the target quantity collected by the experimental device is used for real-time operation to continuously correct the parameters applied to the liquid crystal modulator; generally, the number of training samples collected in the first few cycles is not enough for the neural network algorithm to search for the optimal target value, so the original old parameters will be maintained, and new target values and parameters will be obtained from the experimental system to add training samples. Generally, after about 20 cycles, the search result of the neural network begins to converge gradually, and finally the optimal value is predicted. After that, the algorithm is not closed, and it will always monitor whether the target value (the difference between the output signal of the gyroscope interference and the set value) is greater than the set threshold value. From Figure 2 It can be seen that when the neural network does not work, the proportional integral differential module will always correct the control parameters in a small range in real time. In this way, the experimental system is locked near the optimal working point for a long time, that is, the point at which the polarization coupling error of the experimental system is the smallest. This is the specific process of hybrid circulation.

[0033] The specific role of the proportional integral differential algorithm is to continue small-range fast operation near the polarization locking optimal point of the fiber-optic gyroscope predicted by the neural network algorithm, correct the polarization error and intensity slow drift of the incident light in a long time process, and compensate for the shortcomings of slow feedback speed and long training time of the single neural network algorithm. It is integrated with the neural network algorithm in the python code and belongs to the program part in the present application. As shown in Figure 2 , in the present application, the set classification condition is: Q . Among them, the difference between the output signal of the gyroscope interference and the reference light intensity Q Q. Q , the neural network algorithm will start to run, according to the loss function criterion to find the optimal value, when Q , then directly through the proportional integral differential algorithm for polarization locking. In the present application, the loss function is set as , The current time, represents the gyroscope interference output signal at time t. When the loss function starts to converge continuously, reaches a stable minimum value, the neural network algorithm ends the search, and the optimal polarization locking point of the experimental system can be predicted, and the optimal parameters are transmitted to the proportional integral differential algorithm. The latter will carry out small range fast polarization locking around the optimal parameters. Figure 2 The end condition in the loss function L reaches a stable minimum value (20 consecutive iteration processes cannot find a smaller L value). Thanks to the setting of the experimental hardware system in the present application, the time of a complete iteration process from data collection to algorithm operation, parameter correction, experimental system response, and new data collection is only about 0.1 second.

[0034] The machine learning program used in the present application is the feedforward neural network FNN in the open source package Tensorflow of python. The hybrid machine learning algorithm in the present application is based on the FNN neural network, uses the built-in differential evolution algorithm to generate an initial parameter set, uses the mean square error loss function as the optimization criterion, and is combined with the proportional integral differential PID program according to the process shown in the accompanying Figure 2 . This unique hybrid structure plays a core role in the actual polarization locking process of the fiber optic gyroscope. The single neural network algorithm has practical applications in some physical frontiers such as NV spin reading, light pulse prediction, topological invariant finding, quantum gas preparation, and multi-parameter estimation [1-6] , and has been developed maturely in the fields of biology, chemistry and computer science. At the same time, some hybrid algorithms have also appeared [7-9] . The hybrid algorithm in the present application adopts the framework in the MLOOP open source program package

[10] . This hybrid framework has been used in cold atom BEC experiments, magnetometer coil design, quantum state auxiliary identification, etc. [10-12] , and the use of similar frameworks in the locking of fiber optic gyroscopes is the first time. The present application independently designs the fiber optic gyroscope polarization locking experimental system as shown in the accompanying Figure 1 , and according to the specific experimental parameters and optimization targets, designs a hybrid algorithm program based on the interface of the MLOOP framework. The corresponding algorithm flowchart is shown in the accompanying Figure 2 .

[0035] This invention also proposes a machine learning-based method for polarization locking and drift compensation of fiber optic gyroscopes, utilizing the aforementioned fiber optic gyroscope polarization locking and drift compensation system. In this method, a stable linearly polarized laser beam is split into a 50:50 ratio and enters the loop of the fiber optic gyroscope, interfering with each subsequent exit. Then, a phase modulator shifts the phase, causing the gyroscope's operating point to move to a specific position. Near the phase point, the signal measured by the photodetector is converted to obtain the angular velocity information loaded on the fiber optic loop. The machine learning-based locking method inputs the error value between the photodetector signal and the set value into a neural network for searching and prediction. After calculation using a proportional-integral-differential algorithm, the error is quickly fed back to the signal generator and liquid crystal modulator, completing one cycle and continuing to the next. This cycle continues continuously during the gyroscope's operation. This method is not only applicable to locking the polarization of the incident light in a fiber optic gyroscope, but can also be used for loop loss compensation and rapid phase locking in fiber optic gyroscopes. It has excellent scalability and only requires changing the optimization target value of the machine learning.

[0036] The hybrid machine learning method used in this invention is combined with the aforementioned fiber optic gyroscope experimental system. It continuously cycles through the following steps: parameter initialization, fiber optic gyroscope data acquisition, error signal extraction, neural network training, proportional-integral-differentiator operation, feedback voltage transmission, and liquid crystal modulator response. This process corrects the polarization of the light field in the fiber optic gyroscope in real time, thereby reducing the long-term slow drift of the signal caused by polarization coupling in the fiber optic loop and improving the long-term stability of the fiber optic gyroscope near the operating point.

[0037] The individual sub-algorithms are not the innovation of this invention. This invention simply integrates a neural network algorithm with the traditional proportional-integral-differential (PID) method, using a hybrid algorithm to specifically implement and solve the polarization locking and slow drift compensation problem of fiber optic gyroscopes. This hybrid algorithm differs from existing algorithms in that it combines neural network algorithms and traditional PID algorithms, overcoming the shortcomings of single algorithms. A single neural network algorithm requires a certain number of samples as a training set and cannot quickly compensate and correct the results of the next time step based on the data results of the previous time step, unlike the PID algorithm; while a single traditional PID algorithm relies on multiple preliminary experiments to obtain empirical parameters for the stable operating point of the fiber optic gyroscope. If the initial parameters are not well selected, the locking effect of the algorithm will be greatly reduced. (See attached...) Figure 2 As shown, this paper adopts this hybrid algorithm and Figure 1The experimental system shown is combined to realize polarization locking of the fiber optic gyroscope, thereby improving long-term stability and measurement sensitivity of the fiber optic gyroscope. When the hybrid algorithm is implemented, it is not simply spliced in series. Instead, the neural network algorithm and the proportional-integral-derivative method are parallel circled to meet the actual experimental needs of the fiber optic gyroscope. There are two main difficulties: first, a suitable classification condition needs to be selected. The classification condition determines the actual experimental effect of the neural network algorithm and the proportional-integral-derivative method. If the classification condition is too broad, the search training time of the neural network algorithm may be too long, resulting in that the experiment cannot realize real-time locking. If the classification condition is too accurate, the feedback of the hybrid algorithm may not be able to pull the working point of the gyroscope back to the locking position once it jumps out of the balance position locking point due to external sudden disturbance. Therefore, the classification condition Q is reasonably set according to the actual polarization locking effect of the experimental device in the application, which can shorten the training time of the neural network, resist environmental interference, and improve the robustness of the fiber optic gyroscope when operating near the working point. Second, the neural network algorithm needs to give the optimized best parameters to the PID algorithm for operation. In order to ensure that the best parameters can reflect the best position of the experimental system, the parameters that meet the experimental conditions for optimization need to be selected according to the actual situation of the experiment. For example, there is a mapping relationship between the output light intensity of the fiber optic gyroscope and the voltage of the liquid crystal modulator. Therefore, the voltage and frequency of the liquid crystal modulator are selected as the optimization parameters to ensure this point.

[0038] The specific method includes the following steps:

[0039] Step 1: Run the laser, adjust the optical path, scan the phase modulator to stabilize the interference signal of the fiber optic gyroscope, then obtain the stable interference signal from the photodetector, and then turn off the scanning voltage, use a constant direct current voltage to drive the phase modulator, and fix it at the working phase point of the gyroscope;

[0040] Step 2: Machine learning parameter initialization. By setting the driving voltage amplitude value and frequency of the liquid crystal modulator as optimization parameters, setting the maximum number of iterations and parameter boundary values, generating initial training set data according to the response curve of the liquid crystal modulator, and generating parameters once, the hardware system also responds and starts to collect the next moment of the aforementioned photodetector detection data.

[0041] Step three: On the basis of the initial training set, the neural network algorithm in machine learning starts to predict and build the relationship between the parameters (liquid crystal modulator voltage amplitude value, frequency and other parameters) and the optimization target (the difference between the interferometer output signal of the gyroscope and the set value), and accelerates the convergence process through the built-in L-BFGS and other quasi-Newton iterative algorithms to predict the optimal value and the corresponding parameters. The corresponding parameters will be quickly calculated by the proportional-integral-derivative module and then transmitted to the liquid crystal modulator through the signal generator to correct the previous parameters.

[0042] Step four: After the neural network predicts the optimal parameters, the search process does not end. When the system meets the corresponding classification conditions, the traditional proportional-integral-derivative algorithm will compensate and polarize lock according to the optimal parameters, and repeatedly collect and feedback with the hardware system, and continuously cycle… During this period, the data of the fiber-optic gyroscope collected by the photoelectric detector can still be obtained. The algorithm will determine whether the interference signal has deviated from the optimal value of the previous fiber polarization lock under the interference of external environmental temperature, vibration, stress distortion and other factors, and decide whether to perform a new round of machine learning search and iteration; if it deviates greatly from the optimal value of the previous fiber polarization lock, a new round of machine learning search and iteration is performed.

[0043] The machine learning neural network algorithm in the present application still uses the architecture of the feedforward neural network algorithm, except that the PID proportional-integral-derivative algorithm is integrated into the iterative process of the neural network search according to the actual long-time continuous polarization lock requirement of the fiber-optic gyroscope, and the iterative cycle is performed (mentioned above). The hybrid method has better polarization lock and slow drift suppression effect than the single algorithm. When the fiber-optic gyroscope measures the small angular velocity, the data needs to be accumulated and analyzed by Allan variance noise to improve the angular velocity measurement accuracy. Therefore, the better the polarization lock effect of the fiber-optic gyroscope, the better the long-time stability, and the sensitivity is improved.

[0044] Step five: The fiber-optic gyroscope interference data collected in continuous time is converted into phase change, and then converted into angular velocity change according to the corresponding formula of Sagnac effect. The corresponding angular random walk coefficient, zero bias stability, quantization noise and other working indexes of the gyroscope can be obtained by interval noise analysis by Allan variance. The above indexes are compared with the indexes of the fiber-optic gyroscope locked by the traditional proportional-integral-derivative method and the fiber-optic gyroscope without locking.

[0045] Sagnac effect is a phenomenon describing the phase difference of light propagating along a closed track (in the present application, it refers to the fiber loop) on a rotating platform, and its formula can be expressed as:

[0046] Δφ = 8π × A × Ω / cλ

[0047] where Δφ is the phase difference of light propagation, A is the area enclosed by the fiber loop, Ω is the angular velocity of the rotating platform, λ is the wavelength of light, and c is the propagation speed of light in vacuum.

[0048] Allan variance is a statistical method commonly used to evaluate frequency stability. It can be obtained by sampling and processing the signal in the time domain to get a numerical indicator describing the frequency stability.

[0049] The calculation formula of Allan variance is as follows:

[0050] σ(τ)² = 1 / (2(N-1)) × ∑[y(n+1)-y(n)- <y>]²

[0051] where σ(τ)2is the Allan variance, τ is the sampling interval, N is the total number of sample points, and y(n) is the value of the nth sample point, <y>It is the average value of the entire signal.

[0052] The beneficial effects of this invention compared to the prior art are as follows:

[0053] This invention utilizes machine learning, proportional-integral-differential (PID) and other computational methods to rapidly calibrate polarization, significantly reducing the noise and slow drift of the fiber optic gyroscope's interferometric output signal compared to its unlocked state. In the experiment, the polarization-maintaining fiber loop length of the fiber optic gyroscope was 300m, and the output optical power was 40kJ / m². W. Compared to before polarization locking, the maximum peak-to-peak value of the output signal intensity slow drift has decreased from... Reduced to .

[0054] By introducing machine learning technology, the search capability and resistance to environmental interference of the fiber optic gyroscope are greatly enhanced. Even after being deviated from its original optimal locking position due to external influences, it can still adaptively search and regain the optimal locking parameters. The system's robustness is significantly improved. Under long-term continuous acquisition conditions, thanks to the aforementioned adaptive gyroscope locking method, the corresponding operating indicators such as angle random walk and zero-bias stability are optimized, and the system's measurement sensitivity near the working phase point of the fiber optic gyroscope is improved. In this invention, Allan variance analysis was performed on data collected continuously for 3 hours, and the results are shown in the appendix. Figure 4 and 5 As shown: ARW represents the angular random walk noise of the fiber optic gyroscope, which is also an indicator of its minimum angular velocity sensitivity. BI is an abbreviation for zero-bias instability. Under the same optical power and the same fiber loop length, the measured performance after locking using the traditional PID algorithm is: Angular random walk: 0.0162 Zero-bias instability: 0.133 However, after using a hybrid machine learning approach with feedback locking: the angle random walk result was 0.0131. Zero-bias instability: 0.065 By comparison, it can be found that the zero-bias instability decreased by 0.068. This reflects a significant reduction in intensity drift, indicating a substantial improvement in the long-term stability of the experimental system. Furthermore, the angular random walk coefficient also decreased slightly, reflecting an improvement in the random white noise mixed in with the fiber optic gyroscope output signal after using a hybrid machine learning method. This, in turn, enhances the angular velocity sensitivity of the experimental system under long-term cumulative measurements. More importantly, the system's zero-bias instability index decreased by 0.068. This means that the fiber optic gyroscope's resistance to environmental interference and noise is improved under continuous long-term operation, and the slow drift and zero drift errors during long-term operation are reduced. This is obviously of practical significance for gyroscope measurement. Attached Figure Description

[0055] Figure 1 This is a structural diagram of the machine learning-assisted fiber optic gyroscope polarization locking and slow drift compensation system in this invention.

[0056] Figure 2 This is a schematic diagram illustrating the workflow of the machine learning-based fiber optic gyroscope polarization locking and slow drift suppression method in this invention.

[0057] Figure 3 This is a diagram of the internal structure of the liquid crystal modulator used in the experimental system of this invention.

[0058] Figure 4 This is an Allan variance result graph of the fiber optic gyroscope angular velocity signal locked using a traditional PID algorithm.

[0059] Figure 5 This is an Allan variance result graph of the fiber optic gyroscope angular velocity signal locked using a hybrid machine learning method in this invention.

[0060] Figure 6 This is an interference signal diagram output by the fiber optic gyroscope system in this invention under alternating angular velocity rotation at 200Hz. Detailed Implementation

[0061] The invention will be further described in detail below with reference to the specific embodiments and accompanying drawings. Except for the contents specifically mentioned below, the processes, conditions, and experimental methods for implementing the invention are all common knowledge and general knowledge in the art, and the invention does not have any particular limitations.

[0062] The application provides a machine learning assisted fiber optic gyroscope polarization locking and slow drift compensation system. The system can intelligently search and predict the relationship between the fiber optic gyroscope interference error signal and the incident light polarization direction by introducing a neural network in machine learning, and feed back the predicted optimal value to the hardware system for compensation, thereby realizing the functions of fiber polarization locking and suppressing the intensity slow drift of the output signal. At the same time, the system combines the machine learning algorithm with the traditional proportional integral derivative feedback method, and has the advantages of both, which can adaptively search for the global minimum value of the fiber optic gyroscope error signal, and can also respond quickly in real time to avoid additional interference to the measured angular velocity signal. Compared with the traditional fiber optic gyroscope, the application significantly suppresses the slow drift of the signal, thereby improving the long-time robustness and measurement sensitivity of the system. The specific experimental device comprises a coherent light source, a Glan laser prism, a first polarization beam splitter, a first mirror, a first Faraday rotator, a first half-wave plate, a liquid crystal modulator, a polarization maintaining optical fiber, a phase modulator, a second Faraday rotator, a second half-wave plate, a second mirror, a third half-wave plate, a second polarization beam splitter, a photodetector, a computer and a signal generator. The application has good application background in angular velocity measurement, temperature sensing, space navigation, distributed sensing, deformation detection and the like, and has important significance for precision measurement and national defense science and technology field.

[0063] Reference Figure 1 The machine learning assisted fiber optic gyroscope polarization locking and slow drift compensation system structure comprises a coherent light source 1, a Glan laser prism 2, a first polarization beam splitter 3, a first mirror 4, a first Faraday rotator 5, a first half-wave plate 6, a liquid crystal modulator 7, a polarization maintaining optical fiber 8, a phase modulator 9, a second Faraday rotator 10, a second half-wave plate 11, a second mirror 12, a third half-wave plate 13, a second polarization beam splitter 14, a photodetector 15, a computer 16 and a signal generator 17.

[0064] The coherent light source 1 is a power-stable linearly polarized light, which is further purified after being filtered by the Glan laser prism 2. After the incident light passes through the first polarization beam splitter 3, it is split into a clockwise propagating light around the loop and a counterclockwise propagating light around the loop in equal proportion. In the case that the optical fiber loop does not rotate, the phase difference is constant because the two beams propagate along the same optical axis in the optical fiber loop for the same distance. However, when the angular velocity is loaded on the optical fiber loop, the phase difference between the two beams will change. The angular velocity information can be converted from the change of the interference signal. The first Faraday rotator 5 and the second Faraday rotator 10 are used to ensure that the two beams are combined and emitted at the port of the first polarization beam splitter 3. After the two beams are combined, the third wave plate 13 and the second polarization beam splitter 14 are used to ensure that the polarization is the same to interfere. When the interference signal is observed from the photodetector 15, the phase modulator 9 is needed to move the phase difference value between the two beams, so that the fiber optic gyroscope can be in the state of polarization locking. The phase point is used to work, and the interference signal is measured. The error amount of the interference signal at different times is converted into phase error, and finally the angular velocity information can be converted. The interference signal change of the fiber optic gyroscope caused by the Sagnac effect is . Where Δφ is the phase difference of light propagation, is the amplitude value of the interference signal, A is the area enclosed by the fiber loop, Ω is the angular velocity of the rotating platform, λ is the wavelength of light, and c is the propagation speed of light in vacuum. Therefore, the size of the angular velocity can be extracted from the interference signal by the above formula.

[0065] When the system is in operation, if the error (variance) of the angular velocity deviates from the set threshold value, the machine learning algorithm will search and predict the optimal parameters in the response range with the voltage amplitude and frequency of the liquid crystal modulator as parameters, and then the parameters will be transmitted to the signal generator 17 after being processed by the proportional-integral-derivative algorithm. The signal generator 17 is an arbitrary waveform pulse generator that can reconstruct digital parameters into analog electrical signals to drive the liquid crystal modulator 7 to compensate and feedback the polarization angle.

[0066] The proportional-integral-derivative algorithm (PID algorithm) is a commonly used control algorithm for automatic control. The PID algorithm is based on the feedback control principle, and continuously adjusts the output signal of the controller to make the output signal of the controlled object gradually tend to the target value.

[0067] The PID algorithm includes three parts: proportional (Proportional), integral (Integral), and derivative (Derivative) controllers. The proportional controller calculates the output signal according to the difference between the current output value of the controlled object and the set value; the integral controller calculates the output signal according to the difference between the historical output value of the controlled object and the set value; the derivative controller calculates the output signal according to the difference between the current output value of the controlled object and the output value at the previous time. The output signals of the three controllers are weighted and summed to obtain the final controller output signal.

[0068] The output signal of the proportional controller is proportional to the deviation of the controlled object, that is, the larger the deviation, the larger the output signal. The output signal of the integral controller is proportional to the sum of the historical deviations of the controlled object, that is, the larger the sum of the historical deviations, the larger the output signal. The output signal of the derivative controller is proportional to the difference between the current deviation of the controlled object and the deviation at the previous time, that is, the larger the deviation change rate, the larger the output signal. By comprehensively considering the output signals of the three controllers, the PID algorithm can quickly and accurately adjust the output value of the controlled object, thereby achieving the purpose of automatic control.

[0069] ​In the experimental system, the controlled object is the fiber-optic gyroscope hardware system. The current output value of the controlled object is the interference signal intensity value of the fiber-optic gyroscope. The set value is the signal intensity value of the fiber-optic gyroscope near the interference balance position (static working point). The controller output signal obtained by the PID algorithm after operation and weighting is the voltage and frequency parameter correction value applied to the liquid crystal modulator.

[0070] The complete flow of the polarization locking and slow drift compensation method for the fiber-optic gyroscope based on machine learning is shown in the figure. Figure 2 First, the optical path is adjusted and the phase point is fixed by the phase modulator. Then, the computer program generates initial parameters in the given parameter range by the differential evolution algorithm, and transmits the initial values to the fiber-optic gyroscope hardware system for initialization. The photoelectric signal of the gyroscope is collected by the computer during continuous operation at the fixed phase point. The program calculates the deviation of the measured value from the set value, and compares it with the set classification threshold. When the deviation is less than the threshold, the deviation is directly transmitted to the traditional proportional-integral-derivative algorithm for processing and quickly fed back to the signal generator and liquid crystal modulator, which real-time compensates the polarization angle of the incident light before the polarization-maintaining fiber port, corrects the polarization direction, and continues the next acquisition and feedback cycle. When the deviation is greater than the threshold, it indicates that the polarization direction of the incident light field has deviated from the transmission axis of the polarization-maintaining fiber, causing obvious polarization coupling and corresponding noise and slow drift of intensity. At this time, the deviation data is transmitted to the neural network. The neural network searches and models the adjustable liquid crystal modulator voltage amplitude, frequency, duty cycle and other parameters, and gives the optimized and predicted optimal parameters and values to the proportional-integral-derivative algorithm for rapid processing and feedback to the liquid crystal modulator hardware system for dynamic compensation to complete a complete cycle, and continue the next cycle until the set end condition is reached. The cycle time of the above-mentioned one cycle can be less than 0.1 second, so the method can quickly compensate and suppress the slow drift error, and thus improve the measurement sensitivity. For some high-frequency errors and noises, the polarization locking and slow drift compensation method for the fiber-optic gyroscope based on machine learning is still applicable, but the hardware with shorter response time needs to be replaced, such as electro-optic modulation crystal. In summary, the above-mentioned method based on machine learning can significantly improve the stability and measurement sensitivity of the fiber-optic gyroscope in long-time continuous operation, and has the universality in complex environments, which will help to improve the application value of the fiber-optic gyroscope in actual measurement environment.

[0071] Differential evolution (DE) is an optimization method based on evolutionary algorithm, which is used to solve the optimal solution of the objective function.

[0072] The differential evolution algorithm generates an initial population of individuals randomly, and iteratively finds better individual solutions, and finally obtains the global optimal solution or the local optimal solution of the objective function. The main idea is to generate new individual solutions through selection and mutation operations, and compare and select them with the existing individual solutions, and then find better solutions.

[0073] In the differential evolution algorithm, each individual solution is represented by a vector, and each dimension represents an independent variable. In each iteration, the differential evolution algorithm randomly selects three different individual solutions, linearly combines them to generate a new individual solution, and compares it with the current individual solution to select the better individual solution. This process is called mutation operation. Then, the new individual solution is subjected to differential operation with the current individual solution to obtain a new individual solution, which is compared with the current individual solution to select the better individual solution and update the population. This process is called selection operation.

[0074] The differential evolution algorithm used in the hybrid machine learning method in the present application is mainly used to generate initial parameters for the liquid crystal modulator of the experimental system. The initial parameters and the corresponding experimental data are used as the initial training set of the neural network algorithm. The differential evolution algorithm is built into the neural network program and can be obtained in the open source code TensorFlow package of python.

[0075] The neural network in the hybrid machine learning method adopts FNN feedforward neural network, which iterates parameters according to the principle of back propagation BP, and uses mean square error loss function as the optimization criterion. According to the actual operation of the fiber optic gyroscope, the specific end condition is set as: no smaller L value is found in 20 consecutive iteration processes. The training set data of the neural network is provided by the fiber optic gyroscope output signal and the voltage parameter of the liquid crystal modulator collected at different times. The initial training set is provided by the differential evolution algorithm.

[0076] The core part of the hybrid machine learning algorithm adopted in the present application is provided by the following python calculation package:

[0077] M-LOOP: machine learning online optimization package, which is an open source code package of Git-hub, providing a basic framework and interface for combining neural network algorithm with traditional proportional integral differential algorithm.

[0078] TensorFlow: Google's machine learning package, used to build the neural network part of the hybrid algorithm in the present application;

[0079] Numpy: python scientific calculation package, used for matrix operation on the data collected by the hardware system.

[0080] Embodiment

[0081] According to the Figure 1 The method in the present application is implemented in a fiber-optic gyroscope in the light path structure shown. In the experiment, the fiber-optic gyroscope is first placed in an experimental environment with a temperature of about 20 The entire experimental light path is fixed on an optical platform. As shown in the accompanying Figure 6 In the present embodiment, the interference signal output by the fiber-optic gyroscope when the 300-meter polarization-maintaining fiber ring is rotated under an applied 200Hz alternating angular velocity is tested. The interference pattern contrast reaches more than 96.0%.

[0082] Then, the rotation of the fiber ring is stopped, and the fiber ring is fixed on the platform. In the present embodiment, a 182.3V direct-current constant voltage is applied to the phase modulator to fix the phase operating point of the fiber-optic gyroscope at the balance position of the interference signal, as shown by the dashed line in the accompanying Figure 6 Near the phase point, the fiber ring is most sensitive to a small angular velocity. Therefore, the operating point of the fiber-optic gyroscope in the present embodiment is selected at the balance position of the interference signal. In order to avoid interference caused by changes in the power of the laser itself, the output light intensity of the fiber-optic gyroscope is controlled at 40 microwatts during the entire experiment. However, due to the influence of unavoidable fiber polarization coupling and changes in the ambient temperature, the actual output light intensity of the fiber-optic gyroscope will slowly drift up and down around 40 microwatts. Therefore, when the fiber-optic gyroscope actually measures a small angular velocity (for example, the angular velocity of the earth's rotation), it must be locked. In the present embodiment, a hybrid machine learning algorithm is used to lock the polarization of the experimental system. During the locking process, the voltage parameter and the frequency parameter of the liquid crystal modulator shown in the accompanying Figure 3 are set to 0-5V and 1-10kHz. Compared to before polarization locking, the maximum peak-to-peak value of the slow drift of the output signal intensity is reduced from to .

[0083] Comparative Embodiment

[0084] In the vicinity of the optimal operating point of the fiber-optic gyroscope described above, the polarization is locked using the existing conventional PID locking method and the hybrid machine learning method in the present application, respectively, so that the signal intensity value of the fiber-optic gyroscope is locked at 40 microwatts for a long time, and the slow drift of the signal is compensated. The principle of polarization locking has been described in detail in the foregoing.

[0085] As shown in the accompanying Figure 4 and the accompanying Figure 5 , the polarization is locked using the conventional PID locking method and the hybrid machine learning method in the present application, respectively, the gyroscope data locked for 3 hours in succession is analyzed by Allan analysis, and a comparison is made. Under the same optical power and the same fiber ring length, the indicators obtained by testing after locking by the conventional PID algorithm are: angle random walk: 0.0162 , zero-bias instability: 0.133 ; while using the hybrid machine learning method to lock: angle random walk: 0.0131 , zero-bias instability: 0.065 It can be found that the zero-bias instability is reduced by 0.068 , and the angular velocity resolution of the experimental system under long-term cumulative measurement is improved Therefore, the slow drift compensation method based on machine learning in the application and the experimental system can effectively help the fiber-optic gyroscope to overcome some bottlenecks in the prior art, so that the long-term stability and angular velocity sensitivity in practical application are improved.

[0086] Reference:

[0087] 1. Qian P, Lin X, Zhou F, et al. Machine-learning-assisted electron-spin readout of nitrogen-vacancy center in diamond[J]. Applied Physics Letters, 2021, 118(8): 084001.

[0088] 2. Lohani S, Knutson E M, Zhang W, et al. Dispersion characterization and pulse prediction with machine learning[J]. OSA Continuum, 2019, 2(12):3438-3445.

[0089] 3. Zhang P, Shen H, Zhai H. Machine learning topological invariants with neural networks[J]. Physical review letters, 2018, 120(6): 066401.

[0090] 4. Barker A J, Style H, Luksch K, et al. Applying machine learning optimization methods to the production of a quantum gas[J]. Machine Learning: Science and Technology, 2020, 1(1): 015007.

[0091] 5. Tranter A D, Slatyer H J, Hush M R, et al. Multiparameter optimisation of a magneto-optical trap using deep learning[J]. Nature communications, 2018, 9(1): 4360.

[0092] 6. Carleo G, Cirac I, Cranmer K, et al. Machine learning and the physical sciences[J]. Reviews of Modern Physics, 2019, 91(4): 045002.

[0093] 7. Psichogios D C, Ungar L H. A hybrid neural network-first principles approach to process modeling[J]. AIChE Journal, 1992, 38(10):1499-1511.

[0094] 8. Shon T, Moon J. A hybrid machine learning approach to network anomaly detection[J]. Information Sciences, 2007, 177(18): 3799-3821.

[0095] 9. O’Driscoll L, Nichols R, Knott P A. A hybrid machine learning algorithm for designing quantum experiments[J]. Quantum Machine Intelligence, 2019, 1: 5-15.

[0096] 10. Wigley P B, Everitt P J, van den Hengel A, et al. Fast machine-learning online optimization of ultra-cold-atom experiments[J]. Scientific reports, 2016, 6(1): 25890.

[0097] 11. O’Driscoll L, Nichols R, Knott P A. A hybrid machine learning algorithm for designing quantum experiments[J]. Quantum Machine Intelligence, 2019, 1: 5-15.

[0098] 12. Chen J, Wu Z, Bao G, et al. Design of coaxial coils using hybrid machine learning[J]. Review of Scientific Instruments, 2021, 92(4): 045103.

[0099] The protection scope of the present application is not limited to the above embodiments. Changes and advantages that can be thought of by those skilled in the art without departing from the spirit and scope of the present application are included in the present application, and are protected by the appended claims.< / y> < / y>

Claims

1. A machine learning-assisted fiber optic gyroscope polarization locking and slow drift compensation system, characterized in that, The system includes: a coherent light source (1), a Glan laser prism (2), a first polarization beam splitter (3), a first reflector (4), a first Faraday rotator (5), a first half-wave plate (6), a liquid crystal modulator (7), a polarization-maintaining fiber (8), a phase modulator (9), a second Faraday rotator (10), a second half-wave plate (11), a second reflector (12), a third half-wave plate (13), a second polarization beam splitter (14), a photodetector (15), a computer (16), and a signal generator (17); wherein: The coherent light source (1) is linearly polarized light generated by a continuous laser, which is injected into the interference loop of the gyroscope to sense the phase difference between the clockwise and counterclockwise loops caused by the Sagnac effect. The Glan laser prism (2) has a high extinction ratio, which is used to improve the linear polarization purity of the incident laser and ensure that the polarization direction of the incident light is adjustable; the high extinction ratio means that the extinction ratio is greater than 1:10000. The first polarization beam splitter (3) is used to split linearly polarized light along the horizontal and vertical polarization directions at a 50:50 ratio; The first reflector (4) is used to change the propagation direction of the light field so that the light path coincides with the center line of the subsequent device; The first Faraday rotator (5) is a non-reciprocal device used to change the polarization direction of the light passing in the forward and reverse directions, so that the polarization directions of the incident light and the outgoing light are orthogonal to each other. The first half-wave plate (6) is used to adjust the polarization direction of linearly polarized light; The liquid crystal modulator (7) is composed of a liquid crystal delay plate and a quarter-wave plate, with the slow axis of the liquid crystal delay plate and the quarter-wave plate forming a 45-degree angle. The polarization-maintaining fiber (8) forms a sagnac ring, which generates the sagnac effect when the linearly polarized incident light propagates in the fiber medium, ensuring that the linearly polarized light incident along the optical axis is stable and the polarization of the outgoing light remains unchanged. The phase modulator (9) can change the refractive index difference between the fast axis and the slow axis under the drive of external voltage, thereby generating a phase difference, which is used for phase shifting of interference signals in the fiber optic gyroscope system. The second Faraday rotator (10) makes the polarization directions of the forward incident light and the reverse outgoing light perpendicular to each other; The second half-wave plate (11) is used to adjust the polarization direction of the linearly polarized light and initially align it with the optical axis of the polarization-maintaining fiber. The second reflector (12) can change the direction of light propagation so that the light is collimated and incident on subsequent devices; The third half-wave plate (13) is used to change the polarization state of the light emitted from the optical fiber loop so that the polarization directions of the clockwise emitted light and the counterclockwise emitted light coincide, satisfying the interference condition. The second polarization beam splitter (14), as an analyzer, works in conjunction with the third half-wave plate (13) to maximize the intensity of light transmitted from the second polarization beam splitter (14). The photodetector (15) contains a photodiode, which uses the photoelectric effect to convert light signals into electrical signals; The computer (16) is used to run the photoelectric signal acquisition program and the machine learning program code, and to control the data transmission; The signal generator (17) is an arbitrary waveform pulse generator that is controlled by a computer to combine the input pulse digital signals into analog electrical signals and transmit them to the liquid crystal modulator (7) via a BNC cable.

2. The system as described in claim 1, characterized in that, The coherent light source (1) has a linewidth of 1.5 kHz and stable optical power; The first polarization beam splitter (3) splits the incident light into two beams at a ratio of 50:50, and the polarizations of the two beams after splitting are perpendicular to each other.

3. The system as described in claim 1, characterized in that, The liquid crystal modulator (7) receives electrical signals from the signal generator (17) and responds to the optical path, so that the polarization of the light field is quickly and in real time corrected, suppressing the slow drift of the output signal caused by polarization factors, and the amplitude of the electrical signals it receives must be controlled within the voltage range of the liquid crystal modulator.

4. The system as described in claim 1, characterized in that, When clockwise incident light and counterclockwise incident light pass through the polarization-maintaining fiber (8), they can sense the phase difference caused by the rotation of the external angular velocity, and calculate the angular velocity data of the fiber optic gyroscope using the formula of the Sagnac effect.

5. The system as described in claim 1, characterized in that, The phase modulator (9) can change the relative phase difference between the clockwise and counterclockwise outgoing light in the fiber optic gyroscope when powered by an external voltage source, and fix the operating point of the fiber optic gyroscope near the π / 2 phase point, that is, near the equilibrium position of the interference signal.

6. The system as described in claim 1, characterized in that, The computer (16) is used to collect the signal collected by the photodetector (15) and run the machine learning optimization program. After being processed by the proportional-integral-differential operation module, the signal is output to the signal generator (17). The machine learning program and the proportional-integral-differential module are mixed and cyclically, and the difference between the gyroscope interference output signal and the set value is used as the optimization criterion. The optimization criterion is whether the difference between the gyroscope interference output signal Is and the reference light intensity Q is greater than 1%.

7. The system as described in claim 6, characterized in that, During the mixed loop process of the machine learning program and the proportional-integral-derivative (PID) module, when the difference between the gyroscope interference output signal acquired by the experimental system and the set value exceeds the set threshold, the gyroscope has deviated from the global optimal value for operation. At this point, the machine learning program in the hybrid algorithm quickly searches within a given range to bring the target value back close to the optimal operating point. The PID part performs real-time calculations on the target quantity acquired by the experimental device to continuously correct the parameters applied to the liquid crystal modulator, thus locking the experimental system near the optimal operating point for a long time.

8. A machine learning-assisted method for polarization locking and slow drift suppression of fiber optic gyroscopes, characterized in that, The method employs the system as described in any one of claims 1-7, and the method includes the following steps: Step 1: Run the laser, adjust the optical path, scan the phase modulator (9) to stabilize the interference signal of the fiber optic gyroscope, then obtain a stable interference signal from the photodetector (15), then turn off the scanning voltage, and use a constant DC voltage to drive the phase modulator (9) to fix it at the working phase point of the gyroscope. Step 2: Initialize machine learning parameters. By setting the driving voltage amplitude and frequency of the liquid crystal modulator (7) as optimization parameters, and setting the maximum number of iterations and parameter boundary values, the initial training set data is generated using the differential evolution algorithm based on the response curve of the liquid crystal modulator. Each time parameters are generated, the hardware system responds accordingly and begins to collect the detection data of the photodetector (15) at the next moment. Step 3: Based on the initial training set, the neural network algorithm in machine learning begins to predict and construct the relationship between parameters and optimization objectives, and accelerates the convergence process through the built-in quasi-Newton iterative algorithm to predict the optimal value and optimal parameters; the optimal parameters will be quickly calculated by the proportional-integral-differential module and then transmitted to the liquid crystal modulator (7) through the aforementioned signal generator (17); Step 4: After the machine learning predicts the optimal parameters, it continues to acquire the fiber optic gyroscope data collected from the photodetector (15), judges whether the interference signal has deviated significantly from the previous optimal value, and decides whether a new round of parameter search and model prediction is needed. Step 5: Convert the interferometric data of the fiber optic gyroscope acquired over continuous time into phase change, then convert it into angular velocity change according to the Sagnac effect formula, and perform interval noise analysis using Allan variance to obtain the gyroscope's performance indicators, including the random walk coefficient at the corresponding angle, zero-bias stability, and quantization noise.

9. The method as described in claim 8, characterized in that, The formula for the Sagnac effect is as follows: in, Ω is the phase difference of light propagation, A is the area of ​​the fiber optic loop, Ω is the angular velocity of the rotating platform, λ is the wavelength of light, and c is the speed of light in a vacuum. The formula for calculating the Allan variance is as follows: s(t) 2 =1 / (2(N-1))*∑[y(n+1)-y(n)- <y> ] 2 ，< / y> Where, σ(τ) 2 Let be the Allan variance, τ be the sampling interval, N be the total number of sampling points, and y(n) be the value of the nth sampling point. <y> It is the average value of the entire angular velocity signal.< / y>

Images