A method and system for temperature compensation of fiber optic gyroscopes

Abstract

This invention discloses a method and system for temperature compensation of fiber optic gyroscopes. It acquires the angular velocity output, temperature, and zero-bias drift value under varying temperature conditions, constructs basic features, temporal memory features, and physical interaction feature vectors, and integrates them into a comprehensive physical information feature vector. This comprehensive vector is used as input, and the zero-bias drift value is used as the training label to train the model until convergence. In the online phase, this comprehensive feature vector is reconstructed in real time and input into the model to predict the zero-bias drift, correcting the real-time output angular velocity. The temporal memory feature introduced in this invention quantifies the thermal accumulation history using statistics from a multi-scale sliding window, eliminating the ambiguity problem of thermal hysteresis mapping and significantly reducing computational load. Simultaneously, the physical interaction feature pre-introduces the nonlinear product term of temperature and rate of change into the model, decoupling linear and nonlinear errors, breaking through the generalization bottleneck of traditional black-box models under small sample conditions, and endowing the model with strong physical interpretability and extrapolation capabilities.

Description

Technical Field

[0001] This invention relates to the field of inertial navigation technology, and in particular to a method and system for temperature compensation of fiber optic gyroscopes. Background Technology

[0002] Fiber optic gyroscopes (FOGs) are optical angular velocity sensors based on the Sagnac effect. They measure rotational angular velocity by detecting the phase difference between two counter-propagating beams in an optical fiber loop. Due to their advantages of having no moving parts, a large dynamic range, and high reliability, they are widely used in aerospace, navigation, unmanned systems, and high-precision inertial navigation. However, in practical applications, the accuracy of FOGs is highly susceptible to changes in ambient temperature. The thermal expansion coefficient and refractive index of the fiber optic coil material change with temperature, leading to thermoelastic stress and non-reciprocal Shupe effects within the coil. These thermal effects introduce unpredictable zero-bias drift errors into the gyroscope output, significantly reducing the accuracy and stability of the navigation system. Furthermore, other components in the FOG system are also affected by temperature changes. For example, the multifunctional integrated Y-waveguide, which not only splits and combines light beams but also functions as a phase modulator, is also prone to zero-bias drift and nonlinear response due to temperature variations, further increasing the complexity of system errors. Therefore, temperature compensation for fiber optic gyroscopes is crucial for improving their high-precision operation in complex environments.

[0003] To eliminate zero-bias drift in fiber optic gyroscopes caused by temperature changes, various temperature compensation schemes have been proposed. These can be categorized into traditional compensation methods and machine learning-based compensation methods. Although existing technologies have improved the temperature performance of fiber optic gyroscopes to some extent, they still suffer from the following significant technical drawbacks: (1) It is difficult to strike a balance between eliminating mapping ambiguity and low computing power consumption: the thermal drift of fiber optic gyroscopes has significant thermal hysteresis characteristics, that is, the current zero-bias drift value depends not only on the current temperature value, but also on the previous history of thermal accumulation (heating or cooling process). For example Figure 1 As shown, due to the internal thermal conduction delay of the fiber optic gyroscope, the zero-bias drift exhibits a hysteresis effect with temperature changes. Specifically, at the same temperature point (e.g., 25°C), the fiber optic gyroscope corresponds to different zero-bias drift values ​​during the heating and cooling phases. This phenomenon is mathematically known as a one-to-many mapping, depending solely on the current temperature. TAs input, the model cannot determine whether it is currently on a heating or cooling path, and cannot accurately identify the current zero-bias drift value of the fiber optic gyroscope, thus leading to prediction errors. Existing static temperature compensation models lack memory and cannot distinguish between heating and cooling states, resulting in large errors in the compensated gyroscope output. More importantly, existing methods often lack optimization for embedded low-level data flow in engineering implementation, and do not consider the storage and transfer overhead (O(N) complexity) of large-scale historical data under high-frequency sampling, making it difficult for temperature compensation algorithms to achieve microsecond-level real-time response on computers or field-programmable gate array (FPGA) chips with limited power consumption and computing power.

[0004] (2) Black-box models lack physical interpretability and have poor generalization and extrapolation capabilities: Existing deep learning methods treat the temperature compensation of fiber optic gyroscopes as a black box, relying solely on statistical data patterns for fitting and ignoring the underlying physical mechanisms. This lack of physical mechanisms means that the temperature compensation model is essentially memorizing training data rather than learning its physical laws. In addition, there are often complex and strong correlations between physical quantities, and traditional models struggle to overcome the parameter instability caused by multicollinearity, resulting in poor model reliability in complex environments. Summary of the Invention

[0005] This invention provides a method and system for temperature compensation of fiber optic gyroscopes, which solves the technical problems of existing technologies, such as difficulty in eliminating mapping ambiguity, high computational complexity, lack of physical interpretability, and poor generalization and extrapolation capabilities.

[0006] This invention provides a method for temperature compensation of a fiber optic gyroscope, comprising: Obtain the current sampling time of the fiber optic gyroscope output. t angular velocity oh raw ( t ), the angular velocity at the current sampling moment oh raw ( t Subtracting the reference value yields the true zero-bias drift value. ; Get the current sampling time t temperature T ( t and rate of temperature change Constructing basic feature vectors ; set up m A set of sliding windows at different time scales W ={ w 1, w 2, …, w k ,…, wm}; Regarding the current moment t According to the set of sliding windows W The first in k window size w k Each sliding window's historical temperature sequence was traced back. T ( t ), T ( t -1), …, T ( t - i ), …, T ( t - w k +1)}, calculate the average temperature for this window scale. m wk and temperature standard deviation s wk, Temporal memory feature vectors are constructed using a multi-scale cascade approach. F memory = [ m w1, s w1, m w2, s w2,…, m wk, s wk,… m wm, s wm ]; Build includes n Physical interaction eigenvectors of order-order nonlinear coupling terms ;in, These are the constant coefficients in the Taylor series expansion; From the basic feature vector F basic Temporal memory feature vector F memory Physical interaction feature vector F physics Generate integrated physical information feature vector x ( t )=[ F basic , F memory , F physics ]; The comprehensive physical information feature vector x (t The true zero-bias drift value is used as input. As training labels y ( t ), construct the training dataset D = {( x ( t ), y ( t ))| t =1,2,3,…, N}; Through the training dataset D The temperature compensation model is iteratively trained until it converges; A real-time integrated physical information feature vector is constructed using real-time data. This real-time integrated physical information feature vector is then input into the trained temperature compensation model to output the predicted zero-bias drift value. The temperature-compensated angular velocity is obtained by subtracting the predicted zero-bias drift value from the real-time angular velocity output by the fiber optic gyroscope.

[0007] Specifically, after acquiring the angular velocity and temperature at the current sampling moment of the fiber optic gyroscope output, the process further includes: Set a length of N The sliding detection window, for the angular velocity output by the fiber optic gyroscope at the current sampling moment. oh t Backtracking adjacent N Construct a detection sequence from historical data points { oh t-N+1 , …, oh t} Calculate the local mean of the data within this window. m h With local standard deviation σ h ; If | oh t – m h |≥3σ h The angular velocity at the current sampling time oh t Remove the values ​​and interpolate using the valid values ​​from the previous time step; The temperature data corresponding to the removed angular velocities are interpolated using the Lagrange interpolation method to obtain the temperature value at the current moment.

[0008] Specifically, the training dataset D The temperature compensation model is iteratively trained until convergence, including: For each decision tree that needs to be built in the random forest regression model Tk Random sampling with replacement is used from the training dataset. D Extract one from the training dataset D Subdatasets of the same size D k ; From the sub-dataset D k The optimal split point of the decision tree is searched by randomly selecting a subset of features in all dimensions of the comprehensive physical information feature vector. The number of decision trees is continuously increased for iterative training until the preset convergence condition is met.

[0009] Specifically, after obtaining the temperature-compensated angular velocity, the process further includes: Calculate the root mean square error of the temperature-compensated gyroscope output within the preset observation window; If continuous K If the root mean square error of the gyroscope output after temperature compensation is greater than or equal to the preset aging threshold, then the observation window is truncated. T obs Use the data within to construct an incremental sample set. D new ={( x i , y i ) | i = 1,2,3,…, N new};in, N new The total number of samples in the incremental sample set. x i The first one within the observation window i The comprehensive physical information feature vector extracted at each sampling time point y i To be consistent with the above x i The corresponding true zero-bias drift value of the fiber optic gyroscope; The incremental sample set D new Each of the x i The inputs are respectively fed into the first random forest regression model. k Decision Tree T k And obtain the decision tree T k The feature vector of this integrated physical information x i Predicted zero bias drift value y pred,i; Through formula The decision tree was calculated. T k The incremental sample set under the current aging state D new Error in predicting zero bias drift value E k ; If the error E k If the value exceeds a set threshold, the corresponding decision tree will be pruned and removed from the random forest regression model. Based on the incremental sample set D new Generate using Bootstrap sampling strategy K pruned A new training subset is used to train... K pruned A brand new decision tree { T new1 , T new2 , …, T new_Kpruned};in, K pruned The total number of decision trees removed during pruning; The K pruned A brand new decision tree is merged with the retained decision tree to generate an updated random forest regression model.

[0010] The present invention also provides a fiber optic gyroscope temperature compensation system, comprising: The module for obtaining the true zero-bias drift value is used to acquire the current sampling time of the fiber optic gyroscope output. t angular velocity oh raw ( t ), the angular velocity at the current sampling moment oh raw ( t Subtracting the reference value yields the true zero-bias drift value. ; The basic feature vector construction module is used to obtain the current sampling time. t temperature T ( t and rate of temperature change Constructing basic feature vectors ; The temporal memory feature vector construction module is used to set up... m A set of sliding windows at different time scales W ={w 1, w 2, …, w k ,…, w m}; Regarding the current moment t According to the set of sliding windows W The first in k window size w k Each sliding window's historical temperature sequence was traced back. T ( t ), T ( t -1), …, T ( t - i ), …, T ( t - w k +1)}, calculate the average temperature for this window scale. m wk and temperature standard deviation s wk, Temporal memory feature vectors are constructed using a multi-scale cascade approach. F memory = [ m w1, s w1, m w2, s w2,…, m wk, s wk,… m wm, s wm ]; The Physical Interaction Feature Vector Construction Module is used to construct features containing... n Physical interaction eigenvectors of order-order nonlinear coupling terms ;in, These are the constant coefficients in the Taylor series expansion; The integrated physical information feature vector generation module is used to generate the basic feature vector. F basic Temporal memory feature vector F memory Physical interaction feature vector F physics Generate integrated physical information feature vector x ( t )=[ F basic, F memory , F physics ]; The training dataset construction module is used to construct the comprehensive physical information feature vector. x ( t The true zero-bias drift value is used as input. As training labels, construct the training dataset. D = {( x 1, y 1), ..., ( x n , y n )};in, x n In the first n A comprehensive physical information feature vector constructed at each sampling time point. y n In the first n The true zero-bias drift value corresponding to each sampling time; The temperature compensation model training module is used to train the model using the training dataset. D The temperature compensation model is iteratively trained until it converges; The first zero-bias drift value prediction module is used to construct a real-time comprehensive physical information feature vector through real-time data, input the real-time comprehensive physical information feature vector into the trained temperature compensation model, and output the predicted zero-bias drift value. The temperature compensation module is used to subtract the predicted zero-bias drift value from the real-time angular velocity output by the fiber optic gyroscope to obtain the temperature-compensated angular velocity.

[0011] Specifically, it also includes: The detection sequence construction module is used to set a length of N The sliding detection window, for the angular velocity output by the fiber optic gyroscope at the current sampling moment. oh t Backtracking adjacent N Construct a detection sequence from historical data points { oh t-N+1 , …, oh t} Calculate the local mean of the data within this window. m h With local standard deviation σ h ; Angular velocity interpolation module, used for... oh t – m h |≥3σ hThe angular velocity at the current sampling time oh t Remove the values ​​and interpolate using the valid values ​​from the previous time step; The temperature interpolation module is used to interpolate the temperature data corresponding to the removed angular velocities using the Lagrange interpolation method to obtain the temperature value at the current moment.

[0012] Specifically, the temperature compensation model training module includes: The subset extraction submodule is used for each decision tree that needs to be built in the random forest regression model. T k Random sampling with replacement is used from the training dataset. D Extract one from the training dataset D Subdatasets of the same size D k ; The decision tree optimal split point search submodule is used to search for the optimal split point from the subset of data. D k The optimal split point of the decision tree is searched by randomly selecting a subset of features in all dimensions of the comprehensive physical information feature vector. The iterative training submodule is used to continuously increase the number of decision trees for iterative training until the preset convergence condition is met.

[0013] Specifically, it also includes: The root mean square error calculation module is used to calculate the root mean square error of the temperature-compensated gyroscope output within a preset observation window. Incremental sample set construction module, used for if continuous K If the root mean square error of the gyroscope output after temperature compensation is greater than or equal to the preset aging threshold, then the observation window is truncated. T obs Use the data within to construct an incremental sample set. D new ={( x i , y i ) | i = 1,2,3,…, N new};in, N new The total number of samples in the incremental sample set. x i The first one within the observation window i The comprehensive physical information feature vector extracted at each sampling time point y i To be consistent with the above xi The corresponding true zero-bias drift value of the fiber optic gyroscope; The second zero-bias drift value prediction module is used to predict the incremental sample set. D new Each of the x i The inputs are respectively fed into the first random forest regression model. k Decision Tree T k And obtain the decision tree T k The feature vector of this integrated physical information x i Predicted zero bias drift value y pred,i ; The error calculation module is used to calculate the error using formulas. The decision tree was calculated. T k The incremental sample set under the current aging state D new Error in predicting zero bias drift value E k ; Decision tree pruning module, used to address the error E k If the value exceeds a set threshold, the corresponding decision tree will be pruned and removed from the random forest regression model. The decision tree generation module is used to generate decision trees based on the incremental sample set. D new Generate using Bootstrap sampling strategy K pruned A new training subset is used to train... K pruned A brand new decision tree { T new1 , T new2 , …, T new_Kpruned};in, K pruned The total number of decision trees removed during pruning; The random forest regression model update module is used to update the... K pruned A brand new decision tree is merged with the retained decision tree to generate an updated random forest regression model.

[0014] One or more technical solutions provided in this invention have at least the following technical effects or advantages: 1. Acquire the raw output angular velocity data, synchronized temperature data, and zero-bias drift value of the fiber optic gyroscope under varying temperature conditions. Then, extract the basic feature vector, temporal memory feature vector, and physical interaction feature vector respectively, and integrate them into a comprehensive physical information feature vector. Use the comprehensive physical information feature vector as input to the temperature compensation model, and use the zero-bias drift value as the training label to train the model until convergence. In the online phase, acquire and update historical temperature sequences in real time, and reconstruct the comprehensive physical information feature vector in real time. Input the pre-trained temperature compensation model to obtain the predicted zero-bias drift value and correct the real-time output angular velocity value of the fiber optic gyroscope. The temporal memory feature vector introduced in this invention quantifies the thermal accumulation history using multi-scale sliding window statistics. It replaces complex recursive calculations with lightweight statistical computations, eliminating the mapping ambiguity problem caused by thermal hysteresis and fundamentally solving the one-to-many mapping problem, significantly reducing the learning difficulty and fitting error of the model. On the other hand, it reduces computational complexity, which allows this high-precision compensation algorithm to be deployed in field-programmable gate arrays (FPGAs) or low-power embedded processors, making it of great engineering application value. This invention also utilizes temperature ( T ) and the rate of temperature change ( First-order and higher-order nonlinear product terms are used to construct physical interaction feature vectors (such as...) (etc.). This feature construction method is equivalent to pre-setting the inherent physical laws of the fiber optic gyroscope (i.e., the nonlinear modulation law of the material properties caused by ambient temperature) into the temperature compensation model, forcing the temperature compensation model to separate the linear and nonlinear errors caused by temperature changes. This white-box feature construction strategy enables the model to still have excellent robustness and extrapolation ability in small-sample engineering scenarios, which is unmatched by conventional black-box algorithms. It fundamentally solves the technical problems of traditional data-driven models being prone to overfitting and poor generalization ability under small-sample data, breaks through the generalization bottleneck of black-box models, and has strong physical interpretability.

[0015] 2. Before modeling, this invention introduces a time alignment mechanism to remove outliers and eliminate time deviations in heterogeneous data, thereby ensuring the accuracy of the input model data. This enables the invention to maintain stable compensation performance when faced with electromagnetic interference or sensor data drift, avoiding a decrease in model prediction accuracy due to data quality issues, and effectively enhancing the robustness of fiber optic gyroscopes operating in noisy environments.

[0016] 3. This invention introduces a hyperparameter adaptive optimization strategy, which can automatically balance the number of decision trees and the proportion of features according to actual computing resources and accuracy requirements. This not only reduces the algorithm's computing power requirements, but also ensures that the model always runs at the optimal balance point between accuracy and computing power, avoiding the problem of insufficient model accuracy caused by improper parameter settings.

[0017] 4. This invention also establishes a state monitoring and incremental update mechanism for the temperature compensation model during its usage cycle. This mechanism can sense the change in the root mean square error of the gyroscope output after temperature compensation and trigger the incremental update mechanism when necessary. This ensures the accuracy of the temperature compensation model for long-term use and solves the industry pain point that the accuracy of traditional static models gradually decreases after long-term use. Attached Figure Description

[0018] Figure 1 This refers to the one-to-many phenomenon caused by the thermal hysteresis effect during the heating and cooling process of fiber optic gyroscopes. Figure 2 This is a schematic diagram of the original output of a fiber optic gyroscope under varying temperature conditions. Figure 3 A flowchart illustrating the overall steps of the fiber optic gyroscope temperature compensation method provided in this embodiment of the invention; Figure 4 This is a schematic diagram showing two heating and cooling processes within a sliding window; Figure 5 This is a schematic diagram illustrating the convergence analysis of the prediction error (i.e., root mean square error) of the zero bias drift value and the number of decision trees in this embodiment. Figure 6 This is a schematic diagram comparing the zero-bias drift value predicted by the temperature compensation model in this embodiment with the original zero-bias drift. Figure 7 This is a schematic diagram comparing the output of the fiber optic gyroscope before and after temperature compensation in this embodiment. Detailed Implementation

[0019] This invention provides a method and system for temperature compensation of fiber optic gyroscopes, which solves the technical problems of existing technologies, such as difficulty in eliminating mapping ambiguity, high computational complexity, lack of physical interpretability, and poor generalization and extrapolation capabilities.

[0020] This embodiment uses two fiber optic gyroscopes manufactured in the same batch, numbered FOG1 and FOG2. The test system constructed in this embodiment mainly consists of three parts: fiber optic gyroscopes, a programmable temperature-controlled turntable chamber, and a computer. The fiber optic gyroscope under test is fixed on the platform of the programmable temperature-controlled turntable chamber. The computer is connected to both the fiber optic gyroscope and the temperature-controlled turntable via a standard serial bus. The computer synchronously controls the temperature of the turntable chamber and collects the raw output angular velocity data of the fiber optic gyroscope and the temperature data from the temperature sensor in real time.

[0021] The fiber optic gyroscope used in this embodiment employs a digital closed-loop modulation / demodulation scheme, with internal hardware connected sequentially along the light transmission path. A superluminescent diode serves as a broadband light source, emitting light that enters a multifunctional integrated optical chip (MIOC, i.e., multifunctional Y-branch). This chip integrates the functions of a polarizer, beam splitter, and phase modulator, splitting the beam into two orthogonally polarized beams that are transmitted to the fiber optic sensing loop. This loop, as the core sensing element, is constructed from polarization-maintaining fiber, with its two ends fused to the output pigtails of the MIOC to form a Sagnac interference closed loop. The interferometric light signal is converted into an electrical signal by a photodetector via a return path, and finally input into a digital signal processing circuit to complete closed-loop demodulation and angular velocity calculation. To verify the effectiveness and generalization ability of the model, both FOG1 and FOG2 underwent the same temperature cycling experiment, covering a temperature range of -30°C to +65°C. The initial temperature started at +35°C and increased to +65°C at a rate of 1°C / minute; it was then maintained at a constant +65°C for a period before decreasing to -30°C; finally, it rose back to the initial temperature of +35°C. Throughout the test cycle, the system's temperature changes were controlled by a programmable temperature-controlled turntable chamber, which simultaneously acquired temperature data and angular velocity data output by the FOG at a sampling frequency of 1 Hz and a corresponding sampling time interval of 1 second.

[0022] To show data details, Figure 2 This diagram illustrates the original output of the fiber optic gyroscope under varying temperature conditions in this embodiment, and also shows the temperature changes. Figure 2 It can be seen that with drastic temperature changes, the output of the fiber optic gyroscope exhibits a large nonlinear drift and violent fluctuations.

[0023] To better understand the technical solution of the present invention, the above technical solution will be described in detail below with reference to the accompanying drawings and specific embodiments. Figure 3 As shown, it consists of 6 steps, from step S1 to step S6. Each step will be described in detail below: Step S1: Synchronous Acquisition and Noise Reduction Preprocessing of Multi-Source Data This step aims to collect the raw output of the fiber optic gyroscope and ambient temperature data, and then clean and time-align the data using filtering and interpolation algorithms to provide a high-quality dataset for subsequent feature space construction. For example... Figure 3 As shown, the specific implementation of step S1 includes steps S110 and S120.

[0024] Step S110: Acquire raw output angular velocity data of fiber optic gyroscope under varying temperature conditions. oh raw ( t ) and synchronized temperature data T ( t), and perform noise reduction preprocessing.

[0025] Specifically, the fiber optic gyroscope under test is fixed in a temperature chamber of a programmable temperature-controlled turntable, and both the gyroscope and the turntable are kept stationary. During the preset continuous heating and cooling cycle in the temperature chamber, the real-time angular velocity output of the fiber optic gyroscope is synchronously acquired by the testing system. oh raw ( t ) and temperature sensor data T ( t ).

[0026] To identify and eliminate transient spike noise caused by electromagnetic interference, after acquiring the angular velocity and temperature of the fiber optic gyroscope output at the current sampling moment, the following steps are also included: Set a length of N The sliding detection window, in which, N This indicates the number of sampling points contained in the sliding window (i.e., the window length). N≥ 100). In this embodiment, the sampling frequency is 1Hz, the corresponding sampling time is 1s, and the window length is set. N =100, then the corresponding sliding window time span is 100s. For the angular velocity of the fiber optic gyroscope output at the current sampling moment... oh t Backtracking adjacent N Construct a detection sequence from historical data points { oh t-N+1 , …, oh t} Calculate the local mean of the data within this window. m h With local standard deviation s h ; Specifically, local mean m h The calculation formula is: Local standard deviation s h The calculation formula is: ;in, oh t-i Moving forward from the present moment i The output data value of the fiber optic gyroscope at each sampling time.

[0027] If | oh t – m h | < 3 s h Angular velocity at the current sampling time oh t Considered a valid value; if | oh t – m h |≥ 3 s h Angular velocity at the current sampling time oh t These are considered outliers and removed, and the valid values ​​from the previous time step are used for interpolation. To fill the data gaps created after outlier removal and eliminate step noise introduced by direct filling, Lagrange interpolation was used to interpolate the temperature data corresponding to the removed angular velocities. This interpolation utilized the adjacent values ​​before and after the outlier points. k Using the temperature data corresponding to each valid measurement as the interpolation base point, a polynomial function is constructed to estimate the temperature value at the current moment. In this embodiment, the following method is used: k =4. When the number of selected points is too small (e.g., 1-3), the polynomial order is low, making it difficult to characterize temperature changes over short timescales, leading to interpolation results deviating from the true trend. When the number of selected points is too large, the polynomial order increases, making it prone to oscillations and sensitive to noise, reducing the stability of the interpolated values. In summary, this embodiment uses 4-point interpolation to ensure local fitting ability while avoiding the instability problems caused by high-order interpolation. This step not only ensures the continuity and smoothness of the time series of collected data but also ensures the accuracy of the interpolation results in the basic feature vector. T ( t ) and angular velocity oh raw ( t Strict alignment at physical time avoids feature misalignment caused by data frame loss.

[0028] Step S120: Based on the noise-anti-processed data, obtain the fiber optic gyroscope at the current sampling time. t True zero bias drift value and the rate of temperature change .

[0029] The effective output data obtained after noise reduction preprocessing in step S110 oh raw ( t Subtract the reference value oh ref To obtain the true zero bias drift value .

[0030] Simultaneously, calculate the rate of temperature change at the current sampling time. The rate of temperature change reflects the degree of change in the current ambient temperature and is the direct driving force of thermal drift. Its calculation formula is: In the formula, T ( t-1) is the temperature value sampled at the previous moment. This represents the sampling time interval.

[0031] II. Step S2: Constructing a multidimensional physical perception feature space like Figure 3 As shown, the specific implementation of step S2 includes steps S130, S140, S150 and S160.

[0032] Step S130: Construct the basic feature vector F basic .

[0033] The temperature calculated based on step S120 T ( t and rate of temperature change The two are then concatenated in parallel to generate a basic feature vector representing the real-time temperature state of the fiber optic gyroscope. .

[0034] Step S140: Construct temporal memory feature vector F memory .

[0035] To address the one-to-many mapping problem caused by thermal hysteresis during the heating and cooling process of fiber optic gyroscopes, this embodiment introduces a multi-scale sliding window mechanism, which extracts the average temperature within the window ( m wk ) and temperature standard deviation ( s wk Constructing temporal memory feature vectors F memory This enables precise identification of the thermal state of fiber optic gyroscopes at the physical level.

[0036] The specific implementation steps are as follows: Set up a... m A set of sliding windows at different time scales W = { w 1, w 2,…, w k ,…, w m},in, w k Indicates the first k The number of historical sampling points contained in each sliding window w k The product of the product and the sampling time interval represents the time span of the sliding window. For example, consider a set of sliding windows. W The chosen value is {3, 5, 7, 10, 15, 20, 30}, meaning that at this time... m =7, w 1 = 3,w 2=5, w 3=7, w 4 = 10 w 5 = 15 w 6 = 20 w 7 = 30. The corresponding time spans are 3 seconds, 5 seconds, 7 seconds, 10 seconds, 15 seconds, 20 seconds, and 30 seconds respectively; for the current moment... t According to the sliding window W Each window scale in w k Each of them traces back the corresponding historical temperature series. T seq ( t , w k ), T seq ( t , w k ) = { T ( t ), T ( t -1), …, T ( t - i ),…, T ( t - w k +1)}, where, T ( t - i ) indicates counting backwards from the current moment. i Temperature observations at the sampling point. For the... k Historical temperature series at various scales T seq ( t , w k Calculate their mean values ​​respectively. m wk with standard deviation s wk Specifically, the mean m wk Indicates the first k The average temperature within each window is calculated using the following formula: Units are °C; standard deviation s wk Indicates the first k The dispersion of temperature fluctuations within a window is calculated using the following formula: The unit is °C; after obtaining the statistics of each sliding window mentioned above, a time-series memory feature vector is constructed using a multi-scale cascade approach. Fmemory That is, according to w k The numerical values ​​are arranged in ascending order, and the calculated mean and standard deviation are concatenated to obtain the temporal memory feature vector. F memory = [ m w1, s w1, m w2, s w2,…, m wk, s wk,… m wm, s wm ], its unit is °C, where, w 1< w 2<…< w k <…< w m .

[0037] Average temperature m wk This reflects the level of heat accumulation within the window and serves as a basis for determining the direction of temperature change. By comparing the average temperatures at different time scales... m wk Temperature compensation models can identify trends in temperature changes. s wk It has two functions: First, it measures the temperature standard deviation in the absence of ambient noise and temperature variations. s wk It reflects the drastic degree of temperature change; its second function is to measure the temperature standard deviation in the presence of environmental noise (such as high-frequency circuit noise or mechanical vibration). s wk As an indicator of environmental noise, it triggers the noise resistance mechanism of the temperature compensation model.

[0038] When temperature standard deviation s 3sWhen the abnormal increase occurs, the model determines that noise exists in the current environment and automatically triggers the noise immunity mechanism. In practical applications, the temperature sensor is usually fixed to the outer shell or internal circuit board of the fiber optic gyroscope. Even if the ambient temperature fluctuates drastically, due to the thermal conductivity characteristics of the fiber optic gyroscope's own encapsulation materials (such as the metal shell and internal curing adhesive), there is a certain time delay in heat transfer. Therefore, the temperature change at the measurement point is a relatively stable and continuous gradual process, without instantaneous numerical abrupt changes. Since the sampling frequency in this embodiment is 1Hz (corresponding to a sampling time of 1s), within the 3s observation window (containing only 3 consecutive sampling points), the temperature change at the measurement point will not change significantly, and its corresponding temperature standard deviation is... s 3s The sampling time is relatively short. However, when a fiber optic gyroscope is subjected to electromagnetic pulses or mechanical vibrations, extreme anomalous sampling points may be introduced at a certain sampling moment. Due to the extremely short timescale of the 3-second sliding window, these extreme anomalous sampling points cannot be diluted by a large amount of normal data, which can lead to... s 3s A surge occurs. Therefore, when the short-cycle temperature standard deviation... s 3s When the temperature abnormally increases, the model determines that the change is not a real temperature change, but is caused by environmental noise. After identifying the noise, the model automatically triggers an anti-noise mechanism, reducing the confidence weight of the input data at the current moment, or relying on long-period features for analysis, thereby avoiding interference from environmental noise on the temperature compensation results.

[0039] This embodiment constructs a complete window sequence including intermediate scales. w The reason why only extreme values ​​of the window (3s and 30s) were not selected (∈ {3, 5, 7, …, 30}) is that the fiber optic gyroscope's fiber optic ring is composed of multiple layers of materials, including optical fiber, curing adhesive, skeleton, and metal shell. Heat conduction from the shell to the inside of the coil must pass through complex dielectric layers, each corresponding to a different heat conduction time constant. A single short window (3s) or long window (30s) can only capture the starting and ending states of heat conduction, while intermediate-scale windows (such as 7s and 15s) act as intermediate characterization points in the heat transfer process, which can completely reconstruct the dynamic process of heat diffusion in each layer of the fiber optic ring. The constructed temporal memory feature vector acts as a buffer for short-term memory. By comprehensively comparing the combination relationship of long and short period statistical features, the current real-time thermal response state of the fiber optic gyroscope can be accurately identified, thereby eliminating the ambiguity of one-to-many mapping.

[0040] To address the issue of multiple temperature rises and falls that may occur within a sliding window observation period, relying solely on the difference within a long window (e.g., 30 seconds) can lead to misjudgments. During the heating and cooling process, calculating only the temperature average within 30 seconds will cause the high-temperature peaks during heating to cancel out the low-temperature troughs during cooling, resulting in a temperature average that is the same as or similar to a stationary temperature series. The model will then incorrectly assume that no temperature change has occurred within the current window. However, in reality, multiple heating and cooling processes occur within the window. To address this, the model employs multi-scale mean (…). m wk ) and standard deviation ( s wk A joint determination will be made.

[0041] Combination Figure 4 To illustrate, the sliding window shown in the figure contains two heating and cooling processes. In fact, regardless of the number of heating and cooling processes within the sliding window, it can be decomposed into a combination of four stages: steady-state heating, high-point cooling, low-point heating, and steady-state cooling.

[0042] The model identifies the current temperature change trajectory by extracting the mean temperature gradient from adjacent scale windows. Figure 4 For example, during the steady-state temperature rise phase, t = 5s to t = 10s short-period temperature average m 3s It quickly follows the temperature rise and presents m 3s > m 5s The positive local gradient; while during the cooling phase at the high point, t = 10s to t At 15s, the temperature drops, and the temperature mean flips. m 3s < m 5s The negative local gradient. Utilizing m 3s and m 5s By understanding this gradient change pattern, the temperature compensation model can accurately analyze whether the current temperature change is an upward or downward process.

[0043] However, simply knowing the direction of temperature change is insufficient to assess the true thermal stress inside the gyroscope. This is because the zero-bias drift caused by heating up from a steady state and heating up after cooling are drastically different from that caused by heating up after cooling down. In the former, heat is conducted unidirectionally, resulting in a smaller temperature gradient, less thermal stress, and a relatively gentle zero-bias drift. In the latter, the gyroscope expands externally due to heat while its interior contracts due to cooling. This alternation of heating and cooling leads to a larger temperature gradient, greater thermal stress, and consequently, a larger and more complex zero-bias drift. Therefore, the model incorporates the medium- to long-term temperature standard deviation (in terms of...). s 10s (For example) This is used as a memory indicator for joint judgment. Figure 4 For example, comparing the temperature rise phase in a stable state ( t ≈ 6s) and the low-point warming stage ( t (≈ 16s), during the steady-state heating phase, since the system was previously in a steady state at 20°C, the temperature standard deviation at this time is... s 10s Smaller; while during the low-point warming phase (such as...) t = 16s), because s 10s The backtracking window ( t = 6s~16s) encompasses a period of thermal shock (heating and then cooling process), resulting in s 10s The temperature is relatively high. Therefore, during the steady warming phase, m 3s > m 5s and s 10s Smaller; during the cooling phase in a stable state, m 3s < m 5s and s 10s Smaller; during the cooling phase at the peak, m 3s < m 5s and s 10 Larger; for the low-point warming phase, there is m 3s > m 5s and s 10s Larger. This analysis is limited to... Figure 4 The sliding window shown contains temperature change curves for two heating and cooling processes. If more heating and cooling processes exist, the model can adopt an adaptive scaling approach. m wk and s wkA joint analysis is conducted to determine the current heating or cooling process.

[0044] In summary, this joint decision-making mechanism solves the problem of multiple heating and cooling cycles within a sliding window. By integrating multi-scale temporal features, the model can identify the current thermal state of the fiber optic gyroscope amidst complex temperature changes, thereby finding a unique and accurate zero-bias drift value.

[0045] Step S150: Constructing a system containing... n Physical interaction eigenvectors of order-order nonlinear coupling terms .in, These are the constant coefficients in the Taylor series expansion; This step will be explained in detail. The zero-bias drift of the fiber optic gyroscope mainly originates from the thermally induced non-reciprocal phase difference (i.e., the Shupe effect) within the fiber optic ring. Changes in the external ambient temperature lead to the aforementioned zero-bias drift error. Furthermore, the faster the external temperature changes, the more significant the resulting thermal error. Based on this principle, this embodiment selects an easily measurable temperature change rate. This serves as an indicator of the current thermal driving intensity.

[0046] However, the physical properties of optical fiber materials (such as thermo-optic coefficient and coefficient of thermal expansion) are not constant, but vary with ambient temperature. T A non-linear change occurs. This means that even when facing the same rate of temperature change... (i.e., thermal shock of the same intensity), if the current ambient temperature of the system T The differences will lead to variations in the zero-bias drift value. This is necessary to accurately characterize ambient temperature at the algorithmic level. T With temperature change rate The nonlinear coupling physical mechanism, in this embodiment at a certain reference ambient temperature T The rate of temperature change The resulting zero-bias drift is affected by ambient temperature. T By performing a Taylor series expansion, the following physical expression is derived: (1) In equation (1), The constant coefficients in the Taylor series expansion represent the sensitivity weights of the underlying physical properties of the optical fiber material (such as thermo-optic coefficient, thermal expansion coefficient, etc.) as they change nonlinearly with temperature. The first term on the right-hand side of equation (1) is the basic term, containing the temperature change rate (…). The first term is the linear term; the second term is a first-order coupling term, which includes temperature. T With the rate of change of temperature ( The product term; the third term is a second-order correction term, containing Tsquared and The product term; followed by higher-order correction terms, including temperature. T The higher power of The product term. Equation (1) clearly reveals that the zero bias drift depends not only on the rate of temperature change. It also depends on the temperature. T First and higher orders of the equation and the rate of change of temperature The product term.

[0047] This embodiment constructs a Taylor series expansion form based on the above-described Taylor series expansion, containing... n Physical interaction eigenvectors of order-order nonlinear coupling terms In the formula, the first... k Each component The corresponding Taylor series k The order correction term is used to represent the inclusion of T of k The effect of the power term on the zero bias drift. This physical interaction eigenvector contains both linear and nonlinear errors in the zero bias drift caused by temperature changes. The linear error refers to the first term on the right-hand side of equation (1). That is, the basic term. The nonlinear error refers to the second term on the right side of the equation (1) and the terms following it, which include the first-order coupling term, the second-order correction term, and several higher-order correction terms. This forces the temperature compensation model to analyze the total zero bias drift into linear error terms and nonlinear error terms, thereby decoupling the two, so that the subsequent model can correctly extrapolate the unlearned data based on physical laws.

[0048] Step S160: Construct a comprehensive physical information feature vector x ( t ), to the basic feature vector F basic Temporal memory feature vector F memory Physical interaction feature vector F physics Cascaded splicing generates integrated physical information feature vectors x ( t )=[ F basic , F memory , F physics ].

[0049] This step will be explained in detail, including the basic feature vector. F basic Temporal memory feature vector F memory Physical interaction feature vector Fphysics The components are concatenated sequentially to generate a comprehensive physical information feature vector. x ( t )=[ F basic , F memory , F physics This vector also contains the current thermal state of the fiber optic gyroscope (as determined by...). F basic Representative), thermal hysteresis effect characteristics (by) F memory (representative) and nonlinear physical coupling mechanisms (by) F physics (Representative), providing complete and physically interpretable input data for subsequent temperature compensation models.

[0050] III. Step S3: Adaptive optimization of model hyperparameters like Figure 3 As shown, the specific implementation process of step S3 includes step S170.

[0051] Step S170: Using K The strategy combines cross-validation and grid search, compares the root mean square error of each group of parameters on the validation set, and locks the optimal configuration of the number of decision trees and the maximum feature selection ratio.

[0052] Before model training, in order to balance the fitting accuracy and computational complexity of the random forest regression model, this embodiment introduces an adaptive optimization mechanism, which uses the comprehensive physical information feature vector constructed in step S2. x ( t The true zero-bias drift value obtained in step S1 is used as input. As training labels, construct the basic training set. C .use K The strategy of combining cross-validation and grid search automatically locks in the optimal solution that balances accuracy and efficiency in a multi-dimensional parameter space. The specific process is as follows: (1) Define the parameter search space: For the random forest regression model, select the two core hyperparameters (i.e., the number of decision trees and the feature selection ratio) that have the most significant impact on generalization ability and computational cost as the optimization objects: Number of decision trees: This parameter determines the ensemble size of the model. Set the search interval to [300, 800] and the step size to 50. If the number of decision trees is less than 300, the model will have difficulty sufficiently smoothing the high-frequency noise of the fiber optic gyroscope; if it is more than 800, it will significantly increase the inference time of the embedded system.

[0053] Feature selection ratio: This parameter determines the model's ability to decouple physical features. The search range for the feature selection ratio is set to [0.4, 0.8], with a step size of 0.1. This range covers the transition from strong randomness (0.4) to strong fitting ability (0.8), aiming to help the model find the optimal balance point and prevent overfitting.

[0054] (2) Perform cross-validation optimization process: The basic training set C Randomly divided into M A mutually exclusive subset (selected in this embodiment) M =5, which is equivalent to 5-fold cross-validation. During the grid search process, each parameter combination in the above parameter space is automatically traversed (e.g., number of decision trees = 600, feature selection ratio = 0.6). For each combination, 5 rounds of training and evaluation are performed, and 4 subsets are selected in turn as the training set and the remaining subset is selected as the validation set. The root mean square error (RMSE) of the zero bias drift value of the model on the validation set is calculated.

[0055] (3) Locking the optimal parameter configuration: Calculate the mean RMSE of the zero bias drift value of each parameter combination in 5 rounds of verification, and use it as the performance score of that parameter combination. Compare the scores of all combinations, and select the parameter combination with the smallest mean RMSE as the final optimal hyperparameter configuration.

[0056] After the optimization process described above, the optimal parameters determined in the hardware environment of this embodiment are: 600 decision trees and a feature selection ratio of 0.6. This result shows that the model can most effectively utilize physical interaction features under this configuration, while achieving the lowest prediction error on the validation set. This mechanism ensures that the model can automatically match the individual differences of different batches of fiber optic gyroscopes, achieving fine-grained configuration of one parameter per gyroscope.

[0057] IV. Step S4: Construction and Training of the Ensemble Learning Model like Figure 3 As shown, step S4 is the construction and training of the ensemble learning model, and the specific implementation process of this step includes step S180.

[0058] Step S180: Combine the physical information feature vector x ( t As input, the true zero-bias drift value As training labels, construct the training dataset. D Using datasets D The temperature compensation model was trained using the optimized hyperparameters from step S3. M PI-RFIterative training is performed until convergence. This step involves iterative training using a guided clustering sampling (Bagging) and random feature subspace strategy to establish a nonlinear mapping model from the physical feature space to the zero-bias drift value.

[0059] The specific implementation process is as follows: (1) Construction of training dataset: In this embodiment, the training data is collected in a variable temperature experimental environment. The fiber optic gyroscope is placed in a temperature chamber in a static state (the input angular velocity is zero), and is heated and cooled in a cycle according to a preset temperature curve (from low temperature to high temperature).

[0060] To construct the input-output dataset required for the model, this embodiment calculates the comprehensive physical information feature vector at each time step. x ( t The actual zero-bias drift value measured at the same time is used as the input to the model. As training labels for the model y ( t ), build and generate containing N Training dataset of samples D Its mathematical expression is D = {( x ( t ), y ( t ))| t =1,2,3,…, N In this embodiment, the following is selected: N =18000.

[0061] (2) Bootstrap aggregating sampling: For each decision tree that needs to be built in the random forest regression model T k ( k = 1, …, K ), using random sampling with replacement from the training dataset D Extract one from the training dataset D Subdatasets of the same size D k Because the sampling is random, 36.8% of the samples will not appear in the subset. D k In this process, this mechanism ensures that each decision tree is trained based on a different data distribution, thereby ensuring that the random noise of the fiber optic gyroscope can be effectively smoothed after model integration, and preventing a single outlier from having a global impact on the model.

[0062] (3) Decoupling of physical features based on feature subspace: in constructing decision trees Tk When dealing with any splitting node, the algorithm no longer traverses all features to search for the optimal splitting point, but instead selects from a subset of the dataset. D k Randomly select a preset proportion from all dimensions of the comprehensive physical information feature vector (such as the feature selection proportion locked in step S3). m A subset of features (i.e., the total number of features) participates in the splitting. Assume that the comprehensive physical information feature vector constructed in this embodiment contains 5 core dimensions (i.e., the total number of features). M =5), including the dominant features most strongly correlated with zero bias drift (such as current temperature). T ) and auxiliary features representing physical characteristics (such as coupling terms representing the Shupe effect) The temperature mean term representing thermal hysteresis. m When the feature selection ratio is selected... m When the feature selection ratio is 0.6, the algorithm for constructing each split node in the decision tree will no longer traverse all 5 features, but will randomly select 5 × 0.6 = 3 features to form a candidate subset for searching the optimal split point. With a feature selection ratio of 0.6, approximately 40% of nodes will not have a dominant feature selected during the split. T At these nodes, in order to ensure prediction accuracy, the decision tree is forced to delve deeper into other features (such as...). or m The nonlinear relationship between the zero-bias drift and the zero-bias drift. This strategy is equivalent to forcing a portion of the decision trees to become physical feature experts or hysteresis effect experts, forcing each decision tree to learn from different physical perspectives.

[0063] The random feature subspace strategy used in this embodiment, when constructing any split node of the decision tree, involves associated features (e.g., ... T and There is a high probability that they will not occur simultaneously. This strategy fundamentally undermines... T and The collinearity forces the model to independently learn the weights of each physical term in the local feature space. This strategy realizes the physical mechanism. F physics With time series information F memory Effective decoupling significantly enhances the model's robustness against multicollinearity.

[0064] (4) Iterative Training and Convergence Determination: The temperature compensation model undergoes iterative training by continuously increasing the number of decision trees until a preset convergence condition is met. In this embodiment, the preset convergence condition is that the number of decision trees reaches a preset convergence threshold (600 trees). This threshold is selected based on the convergence analysis of the prediction error as the number of decision trees changes. Figure 5As shown, when the number of decision trees is small (<100), the root mean square error (RMSE) of the predicted zero-bias drift value is large, indicating that the error is high and fluctuates greatly. When the number of decision trees increases to the range of 500-700, the RMSE value shows a significant long-tail feature and tends to flatten out, indicating that the model training is complete. At this time, training can be stopped to ensure that the model accuracy reaches stability and to avoid wasting computing resources due to too many decision trees.

[0065] (5) Model integration and noise suppression: When all K After all decision trees have grown, training stops. The final prediction model is the arithmetic mean of the predictions from all decision trees, thus obtaining the current prediction. t zero bias drift value at time :

[0066] This averaging mechanism is equivalent to a low-pass filter at the signal processing level. While preserving the low-frequency temperature drift trend, it further suppresses the high-frequency angular random walk noise in the fiber optic gyroscope output, thus achieving the dual effects of drift compensation and noise reduction.

[0067] V. Step S5: Online Deployment and Real-time Compensation of the Model Figure 3 The step S5 shown is the online deployment and real-time compensation of the model. The specific implementation of this step includes steps S190 and S200.

[0068] To accommodate the limited storage resources and high real-time requirements of embedded systems (such as field-programmable gate arrays (FPGAs) or low-power digital signal processors (DSPs)), this embodiment employs steps S190 and S200: Step S190: Construct a real-time integrated physical information feature vector using real-time data. x new ( t ), which will generate real-time integrated physical information feature vectors x new ( t Input into the trained temperature compensation model M PI-RF The output yields the predicted zero-bias drift value. .

[0069] Specifically, S190 includes the following underlying hardware implementation mechanisms: (1) Hardware resource configuration and buffer management: During the online operation phase, in order to efficiently process historical data that needs to be backtracked, a fixed-length first-in-first-out (FIFO) circular buffer is set in the random access memory (RAM). The size of the buffer depends on the set maximum sliding window length (e.g.,w =30s). When the new temperature sampling point T ( t When the data point arrives, the old data point is overwritten by moving the write pointer. The circular overwrite strategy used in this embodiment achieves complete independence between data updates and the sliding window scale, regardless of how the window scale is expanded (i.e., the sliding window scale). w (How large is it?) The system only needs to perform one pointer offset and write operation, and its computation time remains constant (i.e., the time complexity is O(1)). This deterministic low latency characteristic effectively eliminates the insufficient computing power and interruption delay caused by large-scale data movement, thereby strictly guaranteeing the microsecond-level real-time response capability of the fiber optic gyroscope at a sampling frequency of 100Hz or even higher.

[0070] (2) Real-time reconstruction of feature vectors: Using the temperature collected at the current moment and the historical data in the buffer, the feature vectors of the comprehensive physical information are reconstructed in real time. x new ( t To reduce computational latency, this embodiment employs a pipelined computing architecture that first performs parallel computation and then uses a recursive algorithm. Parallel computing: In FPGAs, fundamental feature vectors F basic (temperature T Rate of temperature change Temporal memory feature vector F memory (Average temperature) m Temperature standard deviation s ) and physical interaction feature vector F physics ( The computations are performed in parallel.

[0071] Recursive algorithm: For the average temperature of the sliding window ( m ) and temperature standard deviation ( s The calculation of these two statistics uses an incremental recursive formula, which directly updates the current statistics using the statistics from the previous time step and the shifted-in and shifted-out data. This avoids repeatedly traversing and summing all the data in the window, greatly reducing the amount of computation.

[0072] (3) Fast inference of the solidified model: Load the trained and solidified random forest regression model parameters (including the split feature index, split threshold and leaf node weight of all decision trees) into on-chip memory (Block RAM).

[0073] To address the challenge of mapping the irregular tree topology of the Random Forest algorithm to FPGA linear memory, this embodiment employs a flattened array serialization method. All nodes of each decision tree are traversed in level-order or pre-order order, remapping them into a compact one-dimensional array. Each storage unit (Node) in the array is encoded as a fixed-width binary word. Its high-order bits store the split feature index and split threshold, while its low-order bits store the absolute offset addresses of the left and right child nodes in the array. If it is a leaf node, it stores the corresponding prediction weight value.

[0074] During inference, the FPGA logic only needs to read the next memory cell at the corresponding offset address based on the current comparison result (greater than, equal to, or less than the split threshold) to achieve a logic jump. This memory mapping method based on index addressing transforms complex tree traversal operations into deterministic memory read operations. This method not only eliminates the timing uncertainty caused by pointer jumps but also significantly reduces the memory fragmentation rate of Block RAM. Finally, the reconstructed feature vector is... x new ( t The system takes the input model as input, schedules multiple decision trees in parallel to make logical judgments, quickly addresses the corresponding leaf node, reads the zero-bias drift value stored in the leaf node, and accumulates and averages the outputs of all decision trees to obtain the predicted zero-bias drift value at the current time. .

[0075] Step S200: Convert the real-time angular velocity output by the fiber optic gyroscope oh raw ( t Subtract the predicted zero bias drift value The temperature-compensated angular velocity is obtained. oh comp ( t ).

[0076] Specifically, in the underlying hardware execution logic, through formulas oh comp ( t ) = oh raw ( t ) - Real-time correction is completed, and this process includes the following hardware synchronization and output mechanisms: (1) Precise timing alignment of the data link: Since the PI-RF model consumes a fixed number of system clock cycles when performing feature reconstruction and decision tree traversal inference, in order to ensure the physical time consistency of the subtraction operation, there is a delay queue composed of shift registers inside the FPGA / DSP to buffer the raw acquired angular velocity data. ohraw ( t The depth of this delay queue is precisely matched to the number of pipeline ticks required for model inference, thus ensuring the extraction of subtraction operations. oh raw ( t ) and the just calculated They belong to the same sampling time.

[0077] (2) Hardware Arithmetic Operations and Data Stream Output: After timing alignment, a hardware adder / subtractor performs subtraction operations using high-precision data formats (such as 32-bit floating-point or high-bit-width fixed-point numbers). Subsequently, the final angular velocity after temperature compensation is... oh comp ( t It is encapsulated as a standard protocol data frame and output to the outside through an on-chip high-speed serial interface (such as RS-422 or SPI) for computer data analysis or real-time transmission to an inertial navigation system (INS) for attitude calculation.

[0078] Experimental results show that, based on the O(1) buffering mechanism, pipelined inference, and precise timing alignment of the underlying hardware, the processing latency of the entire online compensation process (from data acquisition to output correction) can be strictly controlled at the microsecond level, meeting the low latency requirements of high-speed dynamic navigation applications.

[0079] VI. Step S6: Status monitoring and updating of the temperature compensation model during its usage cycle Figure 3 The step S6 shown is the state monitoring and update of the temperature compensation model during its usage cycle. The specific implementation of this step includes step S210.

[0080] Step S210: By monitoring the root mean square error (RMSE) of the output angular velocity after gyroscope compensation, when continuous K When the error exceeds the preset aging threshold, the temperature compensation model is updated.

[0081] Specifically, addressing the issue of decreased prediction accuracy of the original curing temperature compensation model due to device aging (such as fiber ring encapsulation and fiber coating) during long-term use of fiber optic gyroscopes, this embodiment designs an incremental learning mechanism based on a pruning-growth strategy. The specific implementation steps are as follows: (1) Performance status detection and update triggering: When the fiber optic gyroscope is in a stationary carrier or laboratory condition, calculate the root mean square error of the temperature-compensated angular velocity within a preset observation window (i.e., a preset time period); specifically, continuously collect data for a period of observation window. T obs The angular velocity over a period of time (e.g., 5 minutes) is calculated using a temperature compensation model, and the compensated angular velocity is output. oh comp (t ), through formula Calculate the root mean square error (RMSE) of the compensated angular velocity within this window; where, oh ref ( t ) represents the projection component of the Earth's rotation speed onto the sensitive axis.

[0082] If continuous K If the root mean square error of each calculation is greater than or equal to the preset aging threshold (the aging threshold is set to twice the root mean square error of the gyroscope's output angular velocity after temperature compensation model correction within the full operating temperature range when the fiber optic gyroscope leaves the factory), then it is determined that the current model parameters are out of sync, thereby triggering the incremental learning mechanism.

[0083] (2) Incremental sample set D new Construction: When the incremental learning mechanism is triggered (i.e., if continuous...) K After temperature compensation, once the root mean square error of the gyroscope output is greater than or equal to the preset aging threshold, the current observation window is captured. T obs The data within the specified timeframe (using a 5-minute observation window in this embodiment) is used to construct an incremental sample set. D new = {( x i , y i ) | i = 1,2,3,…, N new};in, N new This represents the total number of samples in the incremental sample set. x i For the first observation window i The comprehensive physical information feature vector extracted at each sampling time point y i To and x i The corresponding true zero-bias drift value of the fiber optic gyroscope. Based on the motion state of the fiber optic gyroscope, this embodiment calculates the drift value according to the following two scenarios. y i And construct an incremental sample set D new : Scenario 1: Stationary State. When the vehicle is in a stationary standby state (such as a vehicle turned off, an aircraft undergoing ground self-check, or a ship anchored at a dock), the input angular velocity of the fiber optic gyroscope is the projection component of the Earth's rotation angular velocity onto the gyroscope's sensitive axis, which serves as the reference value. This component is denoted as... oh ref At this moment, the top is in the... iThe raw output angular velocity at each sampling time point oh raw,i ( t ) contains oh ref and the true zero bias drift value y i Therefore, the true zero-bias drift value y i for:

[0084] Scenario 2: Motion State. When a fiber optic gyroscope is fixed to a device such as an angular rate turntable, the turntable can provide a series of known standard input angular velocities, denoted as... oh ref At this time, the input angular velocity of the turntable oh ref This is the reference value. The true zero-bias drift value. y i Original output angular velocity oh raw,i ( t )and oh ref Difference:

[0085] Whether in a static or dynamic state, when obtaining the zero-bias drift value y i Then, it is compared with the comprehensive physical information feature vector extracted at the corresponding time. x i Perform a one-to-one correspondence to construct an incremental sample set: D new = {( x i , y i ) | i = 1,2,3,…, N new The completed incremental sample set. D new The data is stored in a specific partition of non-volatile memory (NVM) and awaits subsequent incremental learning algorithm calls to update the parameters of the PI-RF temperature compensation model.

[0086] (3) Pruning and evaluation of the temperature compensation model: Read all decision trees in the currently fixed PI-RF temperature compensation model and perform incremental sample set D new Each of the x i The inputs are respectively fed into the first... k Decision TreeT k And obtain the decision tree T k The feature vector of this integrated physical information x i Predicted zero bias drift value y pred,i .

[0087] Through formula Calculation yields the first k Decision Tree T k Incremental sample set under current aging state D new Error in predicting zero bias drift E k .

[0088] This embodiment employs a dynamic threshold strategy based on population statistical distribution: calculating the average prediction error of all decision trees in the current forest on the incremental sample set. In the formula K This represents the total number of decision trees in the random forest model. E k For the above-mentioned k The decision trees in their current aging state apply to the incremental sample set D new Predicting the error of zero bias drift; setting a survival threshold. (in, α (This is a tolerance coefficient, selected within the range of 1.2 - 1.5); if the error of a certain decision tree... E k > E keep This indicates that the physical rules learned by the decision tree are no longer applicable to the aging devices, and these rules should be pruned and removed from the forest; if E k ≤ E keep Then it is retained as a decision tree set { T survived This step aims to filter and retain decision trees that maintain high-accuracy predictions even after device aging.

[0089] (4) Growth and recombination of temperature compensation model: based on incremental sample set D new A bootstrap sampling strategy is used to generate... K pruned A new training subset is used for training. K pruned A brand new decision tree { T new1, T new2 , …, T new_Kpruned};in, K pruned The total number of decision trees to be pruned and removed; unlike the original decision trees, these newly generated decision trees are trained entirely on data after device aging, so they can keenly capture and learn the latest physical mapping relationships (such as changes in thermo-optic coefficients caused by the aging of fiber ring potting compound).

[0090] Will K pruned A brand new decision tree is merged with the retained decision trees to generate an updated random forest regression model, which is the set of surviving decision trees that have been selected and retained. T survived} and the newly generated set of decision trees { T new The models are merged to construct an updated random forest regression model. RF updated Its expression is RF updated ={ T survived}∪{ T new This hybrid ensemble strategy, which preserves the robustness of the old decision tree while adapting to changes with the new decision tree, avoids the forgetting problem caused by retraining and endows the model with the adaptive ability to follow changes in the physical properties of aging devices.

[0091] (5) Secure deployment of embedded systems: In embedded systems, the system generates data through offline or online training phases. K pruned A new decision tree is generated to replace the old decision tree that has degraded in performance or been marked as invalid. During system operation, the parameter data of the newly generated decision tree is used to directly overwrite the storage address of the corresponding invalid decision tree in memory (Flash or RAM), thereby completing the model update. After the reorganization is complete, the system updates the model's version number and checksum, indicating that the model has completed an update.

[0092] To prevent system crashes due to model update failures, this embodiment employs a double-buffering mechanism. After the new temperature compensation model completes training in the background, it must pass a consistency check during subsequent testing (ensuring no performance degradation). Once the model validation is successful, an atomic operation is used to switch the model pointer to the new model and record the update log.

[0093] To verify the effectiveness of the incremental learning mechanism, taking FOG1 as an example: its root mean square error (RMSE) of the gyroscope output angular velocity after temperature compensation model correction at the factory is 0.006° / h. Assuming that after 3 years of use, due to aging, the RMSE of the compensated gyroscope output angular velocity rises to 0.018° / h, exceeding the set aging threshold (twice the root mean square error of the gyroscope output angular velocity after temperature compensation model correction within the full operating temperature range at the time of manufacture, i.e., 0.012° / h), then the update program is automatically triggered, collecting data over a 300-second period to construct an incremental sample set. D new Testing revealed that 128 out of the original model's 600 decision trees were... D new The prediction error exceeded the survival threshold (set at 0.009° / h), so these 128 old decision trees were removed, and based on... D new 128 new decision trees were trained to fill in the gaps. After reorganization and correction by the temperature compensation model, the gyroscope output angular velocity RMSE recovered to 0.005° / h, indicating that the system successfully restored the device accuracy with only about 21% of the model parameters updated, effectively verifying the feasibility of this strategy in embedded systems.

[0094] Corresponding to the temperature compensation method described above, this embodiment of the invention also provides a fiber optic gyroscope temperature compensation system, comprising: The module for obtaining the true zero-bias drift value is used to acquire the current sampling time of the fiber optic gyroscope output. t angular velocity oh raw ( t ), the angular velocity at the current sampling moment oh raw ( t Subtracting the reference value yields the true zero-bias drift value. ; The basic feature vector construction module is used to obtain the current sampling time. t temperature T ( t and rate of temperature change Constructing basic feature vectors ; The temporal memory feature vector construction module is used to set up... m A set of sliding windows at different time scales W ={ w 1, w 2, …, w k ,…, w m}; Regarding the current moment t According to the sliding window setW The first in k window size w k Each sliding window's historical temperature sequence was traced back. T ( t ), T ( t -1), …, T ( t - i ), …, T ( t - w k +1)}, calculate the average temperature for this window scale. m wk and temperature standard deviation s wk, Temporal memory feature vectors are constructed using a multi-scale cascade approach. F memory =[ m w1, s w1, m w2, s w2,…, m wk, s wk,… m wm, s wm ]; The Physical Interaction Feature Vector Construction Module is used to construct features containing... n Physical interaction eigenvectors of order-order nonlinear coupling terms ;in, These are the constant coefficients in the Taylor series expansion; The integrated physical information feature vector generation module is used to generate feature vectors from basic feature vectors. F basic Temporal memory feature vector F memory Physical interaction feature vector F physics Generate integrated physical information feature vector x ( t )=[ F basic , F memory , F physics ]; The training dataset building module is used to integrate physical information feature vectors. x ( tAs input, the true zero-bias drift value As training labels, construct the training dataset. D = {( x 1, y 1), ..., ( x n , y n )};in, x n In the first n A comprehensive physical information feature vector constructed at each sampling time point. y n In the first n The true zero-bias drift value corresponding to each sampling time; The temperature compensation model training module is used to train the model using a training dataset. D The temperature compensation model is iteratively trained until it converges; Specifically, the temperature compensation model training module includes: The subset extraction submodule is used for each decision tree that needs to be built in the random forest regression model. T k Random sampling with replacement is used from the training dataset. D Extract one from the training dataset D Subdatasets of the same size D k ; The decision tree optimal split point search submodule is used to search for the optimal split point from a subset of the dataset. D k The optimal split point of the decision tree is searched by randomly selecting a subset of features in all dimensions of the comprehensive physical information feature vector. The iterative training submodule is used to continuously increase the number of decision trees for iterative training until the preset convergence condition is met.

[0095] The first zero-bias drift prediction module is used to construct a real-time comprehensive physical information feature vector through real-time data, input the real-time comprehensive physical information feature vector into the trained temperature compensation model, and output the predicted zero-bias drift value. The temperature compensation module is used to subtract the predicted zero-bias drift value from the real-time angular velocity output by the fiber optic gyroscope to obtain the temperature-compensated angular velocity.

[0096] It should be noted that, in order to identify and eliminate transient spike noise caused by electromagnetic interference, after obtaining the angular velocity and temperature at the current sampling moment of the fiber optic gyroscope output, the following steps are also included: The detection sequence construction module is used to set a length of N The sliding detection window, in which,N This indicates the number of sampling points contained in the sliding window (i.e., the window length). N≥ 100). In this embodiment, the sampling frequency is 1Hz, the corresponding sampling time is 1s, and the window length is set. N =100, then the corresponding sliding window time span is 100s. For the angular velocity of the fiber optic gyroscope output at the current sampling moment... oh t Backtracking adjacent N Construct a detection sequence from historical data points { oh t-N+1 , …, oh t} Calculate the local mean of the data within this window. m h With local standard deviation s h ; Angular velocity interpolation module, used for... oh t – m h |≥ 3 s h Angular velocity at the current sampling time oh t These are considered outliers and removed, and the valid values ​​from the previous time step are used for interpolation. The temperature interpolation module is used to interpolate the temperature data corresponding to the removed angular velocities using the Lagrange interpolation method, utilizing the adjacent values ​​before and after the outlier. k Using the temperature data corresponding to each valid measurement as the interpolation base point, a polynomial function is constructed to estimate the temperature value at the current moment.

[0097] It should also be noted that, in order to achieve state monitoring and updating of the temperature compensation model throughout its lifecycle, the following is also included: The root mean square error calculation module is used to calculate the root mean square error of the temperature-compensated gyroscope output within a preset observation window. Incremental sample set construction module, used for if continuous K If the root mean square error of the gyroscope output is greater than or equal to the preset aging threshold after temperature compensation, then the observation window is truncated. T obs Use the data within to construct an incremental sample set. D new ={( x i , y i )| i = 1,2,3,…, N new};in, Nnew This represents the total number of samples in the incremental sample set. x i For the first observation window i The comprehensive physical information feature vector extracted at each sampling time point y i To and x i The corresponding true zero-bias drift value of the fiber optic gyroscope; The second zero-bias drift value prediction module is used to predict the incremental sample set. D new Each of the x i The inputs are respectively fed into the first... k Decision Tree T k And obtain the decision tree T k The feature vector of this integrated physical information x i Predicted zero bias drift value y pred,i ; The error calculation module is used to calculate the error using formulas. The decision tree was calculated. T k Incremental sample set under current aging state D new Error in predicting zero bias drift value E k ; The decision tree pruning module is used to address minor errors. E k If the value exceeds a set threshold, the corresponding decision tree will be pruned and removed from the random forest regression model. Decision tree generation module, used for generating decision trees based on incremental sample sets. D new Generate using Bootstrap sampling strategy K pruned A new training subset is used for training. K pruned A brand new decision tree { T new1 , T new2 , …, T new_Kpruned};in, K pruned The total number of decision trees removed during pruning; The random forest regression model update module is used to update the random forest regression model. K prunedA brand new decision tree is merged with the retained decision tree to generate an updated random forest regression model.

[0098] Finally, taking FOG1 as an example, this embodiment compares the zero-bias drift prediction results of two models: the standard RF (Random Forest algorithm) and the PI-RF (Physical Information Ensemble Learning Random Forest algorithm) proposed in this embodiment. Figure 6 (as shown) and the compensation effect after temperature compensation (as shown) Figure 7 (As shown).

[0099] Figure 6 The graph shows a comparison between the original zero-bias drift data of FOG1 after 100s smoothing and the predicted values ​​of the error model. As can be seen from the graph, as the temperature (black solid line) continues to decrease, the original zero-bias drift (blue solid line) exhibits complex nonlinear fluctuations. Comparing the prediction performance of the two models, the standard RF model (green dotted line), while roughly following the downward trend of the drift, shows significant deviations at local peaks and troughs. In contrast, the PI-RF model (purple solid line) demonstrates extremely high fitting accuracy; its predicted trajectory almost perfectly overlaps with the original drift data, accurately capturing not only the overall trend but also keenly reproducing the subtle fluctuations during the drift process.

[0100] Figure 7 The results show a comparison of the fiber optic gyroscope outputs after compensation using two different models. As can be seen from the figure, the output of the uncompensated original gyroscope (blue solid line) is significantly affected by temperature changes, fluctuating wildly within a wide range of 0.9 to 1.35° / h. While the output compensated using the standard RF model (green dotted line) suppresses large drifts to some extent, significant oscillations and instabilities still exist. In contrast, the output compensated using the PI-RF model (purple solid line) exhibits excellent stability, remaining almost entirely on a straight line of approximately 1.21° / h with minimal fluctuations. Analysis of the experimental data shows that the zero-bias stability of FOG1 before compensation is 0.282° / h. After compensation using the standard RF model, this value is optimized to 0.050° / h, while after compensation using the PI-RF model proposed in this embodiment, this value is reduced to 0.006° / h, representing a performance improvement of 97.9%.

[0101] To further verify the generalization ability of the model across different gyroscopes, the same experimental test was conducted on another fiber optic gyroscope, FOG2, in this embodiment. The experimental results show that the zero-bias stability of FOG2 before compensation was 0.146 ° / h, which decreased to 0.046 ° / h after compensation using the standard RF model. However, after compensation using the PI-RF model proposed in this embodiment, this index decreased to 0.005 ° / h, representing a performance improvement of 96.6%. The experimental results of the two fiber optic gyroscopes demonstrate that the average compensation effect of the PI-RF model proposed in this embodiment is as high as 97.2%, fully proving its application value in the research of temperature compensation for fiber optic gyroscopes.

[0102] In summary, to address the problems in existing technologies such as the inability to eliminate one-to-many mapping ambiguity caused by thermal hysteresis, poor physical interpretability, high computational cost and overfitting of deep learning models, lack of robustness against noise interference, and inability of temperature compensation models to cope with the decline in model accuracy due to long-term device aging, this invention provides a method and system for temperature compensation of fiber optic gyroscopes. By introducing a hyperparameter adaptive optimization mechanism, a comprehensive physical information feature vector is constructed, including a basic feature vector, a temporal memory feature vector, and a physical interaction feature vector. Combined with lightweight embedded deployment and an incremental learning mechanism based on the usage cycle of the temperature compensation model, high-precision, full-cycle compensation for thermal drift of the fiber optic gyroscope is achieved. Specifically, this invention first performs multi-level noise-resistant preprocessing on the collected multi-source data and temporally aligns the heterogeneous data; then, it extracts the basic feature vector, temporal memory feature vector, and physical interaction feature vector respectively and performs cascaded splicing to construct a multi-dimensional physical perception feature space, deeply integrating temporal history and physical mechanisms. A hyperparameter adaptive optimization and guided aggregation sampling integration strategy are introduced to establish a PI-RF temperature compensation model. In the online phase, embedded microsecond-level real-time compensation is achieved using a circular buffer with O(1) time complexity. By monitoring the root mean square error of the output angular velocity after compensation during the usage cycle of the temperature compensation model, incremental updates of the model are triggered. This effectively solves the problems of thermal hysteresis mapping ambiguity, difficulty in distinguishing between linear and nonlinear errors, and failure of the temperature compensation model due to device aging. It improves the zero-bias stability, engineering robustness, and long-term accuracy of the fiber optic gyroscope in complex temperature environments. This invention can be applied to inertial navigation systems such as navigation computers, FPGAs, or embedded processors.

[0103] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0104] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure one One or more processes and / or boxes Figure one A device that provides the functions specified in one or more boxes.

[0105] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure one One or more processes and / or boxes Figure one The function specified in one or more boxes.

[0106] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure one One or more processes and / or boxes Figure one The steps of the function specified in one or more boxes.

[0107] Any aspects of this invention not described in detail in the embodiments are well-known techniques to those skilled in the art. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this invention and not to limit it. Although this invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this invention without departing from the spirit and scope of this invention, and all such modifications and substitutions should be covered within the scope of the claims of this invention.

Claims

1. A method for temperature compensation of a fiber optic gyroscope, characterized in that, include: Obtain the current sampling time of the fiber optic gyroscope output. t angular velocity ω raw ( t ), the angular velocity at the current sampling moment ω raw ( t Subtracting the reference value yields the true zero-bias drift value. ; Get the current sampling time t temperature T ( t and rate of temperature change Constructing basic feature vectors ; set up m A set of sliding windows at different time scales W ={ w 1, w 2, …, w k ,…, w m }; Regarding the current moment t According to the set of sliding windows W The first in k window size w k Each sliding window's historical temperature sequence was traced back. T ( t ), T ( t -1), …, T ( t - i ), …, T ( t - w k +1)}, calculate the average temperature for this window scale. μ wk and temperature standard deviation σ wk, Temporal memory feature vectors are constructed using a multi-scale cascade approach. F memory = [ μ w1, σ w1, μ w2, σ w2,…, μ wk, σ wk,… μ wm, σ wm ]; Build includes n Physical interaction eigenvectors of order-order nonlinear coupling terms ;in, These are the constant coefficients in the Taylor series expansion; From the basic feature vector F basic Temporal memory feature vector F memory Physical interaction feature vector F physics Generate integrated physical information feature vector x ( t )=[ F basic , F memory , F physics ]; The comprehensive physical information feature vector x ( t The true zero-bias drift value is used as input. As training labels y ( t ), construct the training dataset D = {( x ( t ), y ( t ))| t =1,2,3,…, N }; Through the training dataset D The temperature compensation model is iteratively trained until it converges; A real-time integrated physical information feature vector is constructed using real-time data. This real-time integrated physical information feature vector is then input into the trained temperature compensation model to output the predicted zero-bias drift value. The temperature-compensated angular velocity is obtained by subtracting the predicted zero-bias drift value from the real-time angular velocity output by the fiber optic gyroscope.

2. The fiber optic gyroscope temperature compensation method as described in claim 1, characterized in that, After acquiring the angular velocity and temperature at the current sampling moment of the fiber optic gyroscope output, the method further includes: Set a length of N The sliding detection window, for the angular velocity output by the fiber optic gyroscope at the current sampling moment. ω t Backtracking adjacent N Construct a detection sequence from historical data points { ω t-N+1 , …, ω t } Calculate the local mean of the data within this window. μ h With local standard deviation σ h ; If | ω t – μ h |≥3σ h The angular velocity at the current sampling time ω t Remove the values ​​and interpolate using the valid values ​​from the previous time step; The temperature data corresponding to the removed angular velocities are interpolated using the Lagrange interpolation method to obtain the temperature value at the current moment.

3. The fiber optic gyroscope temperature compensation method as described in claim 1, characterized in that, The training dataset D The temperature compensation model is iteratively trained until convergence, including: For each decision tree that needs to be built in the random forest regression model T k Random sampling with replacement is used from the training dataset. D Extract one from the training dataset D Subdatasets of the same size D k ; From the sub-dataset D k The optimal split point of the decision tree is searched by randomly selecting a subset of features in all dimensions of the comprehensive physical information feature vector. The number of decision trees is continuously increased for iterative training until the preset convergence condition is met.

4. The fiber optic gyroscope temperature compensation method as described in claim 3, characterized in that, After obtaining the temperature-compensated angular velocity, the method further includes: Calculate the root mean square error of the temperature-compensated gyroscope output within the preset observation window; If continuous K If the root mean square error of the gyroscope output after temperature compensation is greater than or equal to the preset aging threshold, then the observation window is truncated. T obs Use the data within to construct an incremental sample set. D new ={( x i , y i ) | i = 1,2,3,…, N new };in, N new The total number of samples in the incremental sample set. x i The first one within the observation window i The comprehensive physical information feature vector extracted at each sampling time point y i To be consistent with the above x i The corresponding true zero-bias drift value of the fiber optic gyroscope; The incremental sample set D new Each of the x i The inputs are respectively fed into the first random forest regression model. k Decision Tree T k And obtain the decision tree T k The feature vector of this integrated physical information x i Predicted zero bias drift value y pred,i ; Through formula The decision tree was calculated. T k For the incremental sample set under the current aging state D new Error in predicting zero bias drift value E k ; If the error E k If the value exceeds a set threshold, the corresponding decision tree will be pruned and removed from the random forest regression model. Based on the incremental sample set D new Generate using Bootstrap sampling strategy K pruned A new training subset is used to train... K pruned A brand new decision tree { T new1 , T new2 , …, T new_Kpruned };in, K pruned The total number of decision trees removed during pruning; The K pruned A brand new decision tree is merged with the retained decision tree to generate an updated random forest regression model.

5. A fiber optic gyroscope temperature compensation system, characterized in that, include: The module for obtaining the true zero-bias drift value is used to acquire the current sampling time of the fiber optic gyroscope output. t angular velocity ω raw ( t ), the angular velocity at the current sampling moment ω raw ( t Subtracting the reference value yields the true zero-bias drift value. ; The basic feature vector construction module is used to obtain the current sampling time. t temperature T ( t and rate of temperature change Constructing basic feature vectors ; The temporal memory feature vector construction module is used to set up... m A set of sliding windows at different time scales W ={ w 1, w 2,…, w k ,…, w m }; Regarding the current moment t According to the set of sliding windows W The first in k window size w k Each sliding window's historical temperature sequence was traced back. T ( t ), T ( t -1), …, T ( t - i ), …, T ( t - w k +1)}, calculate the average temperature for this window scale. μ wk and temperature standard deviation σ wk, Temporal memory feature vectors are constructed using a multi-scale cascade approach. F memory = [ μ w1, σ w1, μ w2, σ w2,…, μ wk, σ wk,… μ wm, σ wm ]; The Physical Interaction Feature Vector Construction Module is used to construct features containing... n Physical interaction eigenvectors of order-order nonlinear coupling terms ;in, These are the constant coefficients in the Taylor series expansion; The integrated physical information feature vector generation module is used to generate the basic feature vector. F basic Temporal memory feature vector F memory Physical interaction feature vector F physics Generate integrated physical information feature vector x ( t )=[ F basic , F memory , F physics ]; The training dataset construction module is used to construct the comprehensive physical information feature vector. x ( t The true zero-bias drift value is used as input. As training labels, construct the training dataset. D = {( x 1, y 1), ..., ( x n , y n )};in, x n In the first n A comprehensive physical information feature vector constructed at each sampling time point. y n In the first n The true zero-bias drift value corresponding to each sampling time; The temperature compensation model training module is used to train the model using the training dataset. D The temperature compensation model is iteratively trained until it converges; The first zero-bias drift value prediction module is used to construct a real-time comprehensive physical information feature vector through real-time data, input the real-time comprehensive physical information feature vector into the trained temperature compensation model, and output the predicted zero-bias drift value. The temperature compensation module is used to subtract the predicted zero-bias drift value from the real-time angular velocity output by the fiber optic gyroscope to obtain the temperature-compensated angular velocity.

6. The fiber optic gyroscope temperature compensation system as described in claim 5, characterized in that, Also includes: The detection sequence construction module is used to set a length of N The sliding detection window, for the angular velocity output by the fiber optic gyroscope at the current sampling moment. ω t Backtracking adjacent N Construct a detection sequence from historical data points { ω t-N+1 , …, ω t } Calculate the local mean of the data within this window. μ h With local standard deviation σ h ; Angular velocity interpolation module, used for... ω t – μ h |≥3σ h The angular velocity at the current sampling time ω t Remove the values ​​and interpolate using the valid values ​​from the previous time step; The temperature interpolation module is used to interpolate the temperature data corresponding to the removed angular velocities using the Lagrange interpolation method to obtain the temperature value at the current moment.

7. The fiber optic gyroscope temperature compensation system as described in claim 5, characterized in that, The temperature compensation model training module includes: The subset extraction submodule is used for each decision tree that needs to be built in the random forest regression model. T k Random sampling with replacement is used from the training dataset. D Extract one from the training dataset D Subdatasets of the same size D k ; The decision tree optimal split point search submodule is used to search for the optimal split point from the subset of data. D k The optimal split point of the decision tree is searched by randomly selecting a subset of features in all dimensions of the comprehensive physical information feature vector. The iterative training submodule is used to continuously increase the number of decision trees for iterative training until the preset convergence condition is met.

8. The fiber optic gyroscope temperature compensation system as described in claim 7, characterized in that, Also includes: The root mean square error calculation module is used to calculate the root mean square error of the temperature-compensated gyroscope output within a preset observation window. Incremental sample set construction module, used for if continuous K If the root mean square error of the gyroscope output after temperature compensation is greater than or equal to the preset aging threshold, then the observation window is truncated. T obs Use the data within to construct an incremental sample set. D new ={( x i , y i )| i = 1,2,3,…, N new };in, N new The total number of samples in the incremental sample set. x i The first one within the observation window i The comprehensive physical information feature vector extracted at each sampling time point y i To be consistent with the above x i The corresponding true zero-bias drift value of the fiber optic gyroscope; The second zero-bias drift value prediction module is used to predict the incremental sample set. D new Each of the x i The inputs are respectively fed into the first random forest regression model. k Decision Tree T k And obtain the decision tree T k The feature vector of this integrated physical information x i Predicted zero bias drift value y pred,i ; The error calculation module is used to calculate the error using formulas. The decision tree was calculated. T k For the incremental sample set under the current aging state D new Error in predicting zero bias drift value E k ; Decision tree pruning module, used to address the error E k If the value exceeds a set threshold, the corresponding decision tree will be pruned and removed from the random forest regression model. The decision tree generation module is used to generate decision trees based on the incremental sample set. D new Generate using Bootstrap sampling strategy K pruned A new training subset is used to train... K pruned A brand new decision tree { T new1 , T new2 , …, T new_Kpruned };in, K pruned The total number of decision trees removed during pruning; The random forest regression model update module is used to update the... K pruned A brand new decision tree is merged with the retained decision tree to generate an updated random forest regression model.

Images