A hybrid model-based method for separating and estimating multi-source errors of MEMS gyroscopes

Through wavelet transform decomposition and hybrid model (ARMA+LSTM) combined with CKF filtering algorithm, the non-stationarity problem of random error in MEMS gyroscope is solved, real-time accurate estimation and compensation of gyroscope random errors is realized, and measurement accuracy is improved.

CN118640937BActive Publication Date: 2024-11-12HANGZHOU INNOVATION RES INST OF BEIJING UNIV OF AERONAUTICS & ASTRONAUTICS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411104539.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-11-12
Estimated Expiration
2044-08-13

AI Technical Summary

Technical Problem

MEMS gyroscopes have large measurement errors in navigation and pose applications, especially random errors, resulting in low accuracy. The existing technologies such as ARMA models and RNN networks have insufficient prediction accuracy and generalization capabilities when dealing with non-stationary errors. A problem of insufficient.

Method used

Wavelet transform is used to extract the trend and detailed information of the MEMS gyroscope error sequence, decompose it into stationary and non-stationary subsequences, and ARMA and LSTM models are established respectively, and real-time estimation and compensation of gyroscope random errors are achieved by combining CKF filtering algorithm.

Benefits of technology

Through the real-time filtering framework of the hybrid model, it is possible to more accurately separate and compensate gyro random errors, improve measurement accuracy, reduce the autocorrelation of residuals, make it close to white noise, and fully extract useful information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118640937B_ABST
    Figure CN118640937B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for separating and estimating multi-source errors of a micro-electromechanical gyroscope based on a hybrid model. First, the output of a micro-electromechanical gyroscope at zero input is collected and preprocessed by mean filtering; secondly, the output data is decomposed by wavelet transform to obtain a stationary subsequence and a non-stationary subsequence; thirdly, an autoregressive moving average ARMA model is established for the stationary subsequence, and a long short-term memory network LSTM model is established for the non-stationary subsequence, and the ARMA model and the LSTM model are combined to obtain a hybrid model of gyro random errors; finally, based on the established hybrid model, a cubature Kalman filter CKF is used to perform real-time separation and estimation of gyro random errors, and the obtained estimated value is used for gyro output compensation. The present invention uses wavelet transform to establish a hybrid model, and designs CKF on the basis of the hybrid model to realize the separation and estimation of gyro random errors, thereby obtaining more accurate sensor information and improving navigation accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the field of navigation sensor noise estimation and relates to a method for separating and estimating multi-source errors of a micro-electromechanical gyroscope based on a hybrid model. The method establishes a hybrid model of random errors of a MEMS gyroscope through wavelet transform and incorporates the hybrid model into the framework of a CKF filtering algorithm to achieve separation and real-time estimation of gyroscope random errors. Background Art

[0002] Currently, MEMS gyroscopes have been widely used in inertial technology fields such as navigation and attitude determination due to their significant advantages such as low cost, small size, and low power consumption. However, due to the large measurement error of MEMS gyroscopes, the long-term orientation accuracy is still not ideal, which restricts their further application. Therefore, it is of great significance to carry out MEMS gyroscope error analysis and separation estimation algorithm design to improve measurement accuracy.

[0003] The measurement errors of MEMS gyroscopes mainly include deterministic errors and random errors. Common deterministic errors include scale factor error, constant error, installation deflection error, and lever arm error, which can usually be compensated by experimental calibration. Random errors caused by complex random interferences such as mechanical noise and electronic noise cannot be eliminated by calibration. Therefore, random errors are the main factor restricting the improvement of MEMS gyroscope accuracy. In order to analyze and reduce random errors through numerical methods, various methods have been proposed, which can be divided into two categories: traditional statistical methods and artificial intelligence methods.

[0004] The paper "Adaptive H∞ Kalman filter based random drift modeling and compensation method for ring laser gyroscope" establishes an ARMA model for random errors, and on this basis, integrates H∞ technology and adaptive ideas into Kalman filtering to estimate the statistical characteristics of unknown noise and realize online compensation of random errors of ring laser gyroscope (RLG). However, the ARMA model simply regards random errors as a linear combination of their previous values ​​and previous driving noise, and has limited extraction of useful information in the sequence. Chinese authorized patent CN 107330149 B establishes an ARMA model based on the original gyro random error data, and uses BP neural network to learn the ARMA model residuals. The combined prediction method of ARMA and BP neural network is used to deal with the weak time-varying and weak nonlinear characteristics of gyro data. However, the premise of establishing the ARMA model is that the sequence is stationary, but the actual gyro random error often exhibits non-stationarity. Therefore, direct ARMA modeling will make the order and parameter selection of the model inaccurate, affecting the prediction accuracy. The paper “Recurrent-neural-network-based unscented Kalman filter for estimating and compensating the random drift of MEMS gyroscopes in real time” uses an RNN network to establish a generalized nonlinear autoregressive moving average (NARMA) model, breaking the traditional ARMA model’s assumption that random errors are a linear combination of their previous values ​​and previous driving noise. However, RNN requires a large amount of data for training. If the training data is insufficient, it may face the problem of insufficient generalization ability, and its accuracy in short-term sequence prediction is not as good as that of the ARMA model. Summary of the invention

[0005] The technical problem to be solved by the present invention is: in view of the adverse effects of multi-source errors of MEMS gyroscope sensors on measurement accuracy, a micro-electromechanical gyroscope multi-source error separation and estimation method based on a hybrid model is proposed, a hybrid model of ARMA and LSTM is established through wavelet transform, and the hybrid model is incorporated into the framework of CKF filtering algorithm to realize real-time estimation and compensation of gyroscope random errors, thereby improving measurement accuracy to a certain extent.

[0006] The technical solution adopted by the present invention to solve the above technical problem is: a method for separating and estimating multi-source errors of a micro-electromechanical gyroscope based on a hybrid model, and the implementation steps are as follows:

[0007] Step 1: Data collection and preprocessing: Collect the output of the MEMS gyro when zero input is applied, and preprocess the MEMS gyro error data. Compensate for the constant deviation in the actual collected error sequence, and use the mean filter algorithm to eliminate the white noise in the sequence data. The filter window is set to 0.1s.

[0008] Step 2: Error sequence decomposition: The random error sequence of the MEMS gyroscope is recorded as , using wavelet transform to decompose and extract trend information and detail information in the sequence. After KPSS stationarity test, the sequence is divided into a stationary subsequences and 1 non-stationary subsequence.

[0009] In order to model and predict the gyro error sequence, wavelet transform is introduced to extract the stationary components in the non-stationary sequence. The specific implementation is as follows:

[0010] The random error sequence of the MEMS gyroscope after mean filtering is recorded as , After mean filtering The observation value of random error at each moment. Select the Daubechies wavelet basis function and use discrete wavelet transform to decompose the non-stationary sequence. When performing the first layer of wavelet decomposition, the original signal is passed through a low-pass filter and a high-pass filter respectively to decompose the original signal into a trend sequence. and detail sequence When the next wavelet decomposition is performed, the trend sequence obtained from the previous decomposition is Further decomposed into trend series and detail sequence Repeat the above process until the set number of decomposition levels is reached. The secondary wavelet decomposition formula is: in, Indicates After sub-wavelet decomposition, the trend sequence is obtained. is the sequence length; For the The trend sequence after decomposition is is the sequence length; For the The detailed sequence after decomposition, is the sequence length; represents a low-pass filter; Represents a high-pass filter.

[0011] Signal reconstruction is performed for each subsequence so that the subsequence length is consistent with the original sequence. The reconstruction formula is as follows: Perform KPSS stationarity test on each subsequence obtained by decomposition, and get stationary subsequences and non-stationary subsequences, among which and are both positive integers greater than or equal to 1, and the sum of the two is the total number of subsequences. It is found from actual data processing that b is often 1. On the other hand, b non-stationary subsequences can be combined into 1 non-stationary subsequence by addition, so it can be considered that wavelet decomposition obtains a stationary subsequence and 1 non-stationary subsequence.

[0012] Step 3: Establish an ARMA model for stationary subsequences: Establish an ARMA random time series model for each stationary subsequence. The specific implementation is as follows: The model is represented as follows: in , are the autoregressive order and moving average order of the ARMA model respectively; Represents the predicted value of the stationary subsequence at time t; is the historical observation value of the stationary subsequence; is the historical random interference sequence; and is the model coefficient; is zero mean white noise. The specific process of establishing the ARMA model is as follows:

[0013] First, use the AIC criterion to Determine the order of the model by plotting the heat map corresponding to the AIC value to intuitively find the order with the smallest AIC value and , considering the complexity of the model, the design and is an integer in the interval [0, 10]. The AIC criterion can be expressed as: in is the maximum likelihood estimate of the model residual variance; is the number of observations.

[0014] After determining the model order, the maximum likelihood estimation method is used to fit the model parameters to obtain the ARMA model.

[0015] Step 4: Establish the LSTM network model of non-stationary subsequences. The specific implementation is as follows:

[0016] In the LSTM calculation process, each LSTM unit has three gates: input gate, forget gate and output gate. The forget gate is used to control the effect of the previous state on the current memory unit state value. Impact: In the formula is the sigmoid activation function; is the input sequence value at time t; and is the weight matrix; is the offset term; The output of the previous moment.

[0017] The input gate is used to control the input of the current time step The state value of the memory unit Impact: In the formula , , and is the weight matrix; and is the offset term; tanh is the activation function; For the current moment Bring new information; It is the memory unit state after combining the current and last updated state.

[0018] The output gate determines the output of the current time step: The output value at time t is After scaling with the tanh function The output at this moment is .

[0019] Non-stationary subsequences for random errors of MEMS gyroscopes The LSTM model is established. The specific implementation steps are as follows: the collected non-stationary sequence data is divided into several samples, and a rolling window with a window size of m is used. The window size refers to the input time period. Each time window constitutes a sample as the input of LSTM, and the sequence value of the next moment of the time window is the target value. After training with the training set, the parameters of LSTM can be obtained, including the weight matrix and the offset term. Therefore, the single-step prediction of the non-stationary subsequence of the gyro random error can be defined as using the non-stationary subsequence values ​​of the current moment and the previous m-1 moments, a total of m data, to predict the non-stationary subsequence value of the next moment, as follows: in, Represents the LSTM model. By training the LSTM network, the dynamic characteristics of the non-stationary subsequence can be learned, and the non-stationary subsequence value at the next moment can be accurately predicted, and the internal state of the network can be continuously updated.

[0020] Step 5: Based on the third and fourth steps, the hybrid model of MEMS gyro random error can be obtained Since the model of random error is a nonlinear function, the cubature Kalman filter CKF is used to realize the real-time estimation and compensation of random error. The specific implementation is as follows:

[0021] According to the third and fourth steps, the hybrid model of MEMS gyroscope random error can be obtained: in, Represents the random error prediction value of the hybrid model at time k; Represents the ARMA model of the ith stationary subsequence; LSTM model representing non-stationary subsequences; in order to compensate for the random error of MEMS gyroscope in real time, the hybrid model is introduced into the filtering framework. Considering that too high a model order will lead to a sharp increase in filtering calculations, the system state is defined according to the network model , is the non-stationary subsequence of the random error of MEMS gyroscope.

[0022] The system state equation and measurement equation can be expressed as: in, Represents the one-step prediction error of the network model. is a nonlinear function, so the state equation is a nonlinear equation; is the observed value of the MEMS gyro under static (zero input) conditions at time k. After compensating the predicted values ​​of all stationary subsequences, the measured value of the non-stationary subsequence containing white noise can be obtained. .

[0023] The system equations can be simplified to: in, ; is the measurement matrix; represents the process noise, The covariance of ; represents a nonlinear transfer function; represents the measurement noise.

[0024] The steps of the MEMS gyroscope multi-source error separation estimation method based on the hybrid model are as follows:

[0025] (1) Prediction stage

[0026] right The optimal estimate of the state at the moment Sampling from the posterior probability distribution: In the formula For the Volume points; and Respectively The posterior state estimate and error covariance matrix at time t; Represents the square root matrix obtained by Cholesky decomposition of the matrix List.

[0027] (2) Time update phase

[0028] The sample points are transformed by nonlinear functions and the predicted mean and covariance are calculated.

[0029] a) The volume points realize state transfer through nonlinear transfer function: The LSTM network in the hybrid model is When transferring volume points, no internal state update is performed. After the state is transferred, the optimal estimate of the state at the previous moment As the input of LSTM, the internal state of LSTM network is updated to be the volume point at the next moment Prepare for the transfer.

[0030] b) Weighted calculation of the prior probability distribution of the state variables at time k: (3) Measurement update phase Therefore, the optimal state estimate at time k is and its error covariance for: in .

[0031] In summary, by establishing a hybrid model for the random error of the MEMS gyroscope and introducing the CKF filtering framework, the optimal prediction of the random error at time k is obtained. .

[0032] In the real-time filtering process, the estimated random error can be used to compensate the raw output of the gyro , the autocorrelation of the residual after compensation is close to the Dirac function, indicating that the useful information in the original sequence has been extracted.

[0033] Compared with the prior art, the present invention has the following advantages:

[0034] (1) In reality, the gyro error sequence is disturbed by random errors and often exhibits non-stationarity. At this time, the accuracy of the ARMA model commonly used for stationary sequence prediction is limited. To address this problem, the present invention proposes the concept of a hybrid model, which uses wavelet transform to extract the stationary component in the non-stationary error sequence to achieve the separation of random errors. ARMA model and LSTM model are established for the stationary subsequence and the non-stationary subsequence respectively, thereby obtaining a hybrid model of MEMS gyro random error. This model can make full use of the useful information in the sequence and lays the foundation for real-time and accurate estimation of random noise.

[0035] (2) The present invention incorporates the hybrid model of gyro random error into the framework of the real-time filtering algorithm CKF and performs appropriate combination design to achieve real-time estimation and compensation of random error. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flow chart of a method for separating and estimating multi-source errors of a micro-electro-mechanical gyroscope based on a hybrid model according to the present invention;

[0037] Figure 2 This is the wavelet decomposition diagram of the random error of MEMS gyroscope;

[0038] Figure 3 The one-step prediction result diagrams for the ARMA model, LSTM network model, and hybrid model;

[0039] Figure 4 Normalized autocorrelograms of the residuals before and after compensation for the three methods ARMA-KF, LSTM-CKF, and hybrid model-CKF. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0041] The method of the present invention is specifically implemented as follows:

[0042] Step 1: Data collection and preprocessing: Collect the output of the MEMS gyro when zero input is applied, and preprocess the MEMS gyro error data. Compensate for the constant deviation in the actual collected error sequence, and use the mean filter algorithm to eliminate the white noise in the sequence data. The filter window is set to 0.1s.

[0043] Step 2: Error sequence decomposition: The random error sequence of the MEMS gyroscope is recorded as , using wavelet transform to decompose and extract trend information and detail information in the sequence. After KPSS stationarity test, the sequence is divided into a stationary subsequences and 1 non-stationary subsequence.

[0044] In order to model and predict the gyro error sequence, wavelet transform is introduced to extract the stationary components in the non-stationary sequence. The specific implementation is as follows:

[0045] The random error sequence of the MEMS gyroscope after mean filtering is recorded as , After mean filtering The observation value of random error at each moment. Select the Daubechies wavelet basis function and use discrete wavelet transform to decompose the non-stationary sequence. When performing the first layer of wavelet decomposition, the original signal is passed through a low-pass filter and a high-pass filter respectively to decompose the original signal into a trend sequence. and detail sequence When the next wavelet decomposition is performed, the trend sequence obtained from the previous decomposition is Further decomposed into trend series and detail sequence Repeat the above process until the set number of decomposition levels is reached. The secondary wavelet decomposition formula is: in, Indicates After sub-wavelet decomposition, the trend sequence is obtained. is the sequence length; For the The trend sequence after decomposition is is the sequence length; For the The detailed sequence after decomposition, is the sequence length; represents a low-pass filter; Represents a high-pass filter.

[0046] Signal reconstruction is performed for each subsequence so that the subsequence length is consistent with the original sequence. The reconstruction formula is as follows: Perform KPSS stationarity test on each subsequence obtained by decomposition, and get stationary subsequences and non-stationary subsequences, among which and are both positive integers greater than or equal to 1, and the sum of the two is the total number of subsequences. It is found from actual data processing that b is often 1. On the other hand, b non-stationary subsequences can be combined into 1 non-stationary subsequence by addition, so it can be considered that wavelet decomposition obtains a stationary subsequence and 1 non-stationary subsequence.

[0047] Step 3: Establish an ARMA model for stationary subsequences: Establish an ARMA random time series model for each stationary subsequence. The specific implementation is as follows: The model is represented as follows: in , are the autoregressive order and moving average order of the ARMA model respectively; Represents the predicted value of the stationary subsequence at time t; is the historical observation value of the stationary subsequence; is the historical random interference sequence; and is the model coefficient; is zero mean white noise. The specific process of establishing the ARMA model is as follows:

[0048] First, use the AIC criterion to Determine the order of the model by plotting the heat map corresponding to the AIC value to intuitively find the order with the smallest AIC value and , considering the complexity of the model, design and is an integer in the interval [0, 10]. The AIC criterion can be expressed as: in, is the maximum likelihood estimate of the model residual variance; is the number of observations.

[0049] After determining the model order, the maximum likelihood estimation method is used to fit the model parameters to obtain the ARMA model.

[0050] Step 4: Establish the LSTM network model of non-stationary subsequences. The specific implementation is as follows:

[0051] During the LSTM calculation process, each LSTM unit has three gates: input gate, forget gate, and output gate.

[0052] The forget gate is used to control the state of the previous moment to the current memory unit state value Impact: In the formula is the sigmoid activation function; is the input sequence value at time t; and is the weight matrix; is the offset term; The output of the previous moment.

[0053] The input gate is used to control the input of the current time step The state value of the memory unit Impact: In the formula , , and is the weight matrix; and is the offset term; tanh is the activation function; For the current moment Bring new information; It is the memory unit state after combining the current and last updated state.

[0054] The output gate determines the output of the current time step: The output value at time t is After scaling with the tanh function The output at this moment is .

[0055] Non-stationary subsequences for random errors of MEMS gyroscopes The LSTM model is established. The specific implementation steps are as follows: the collected non-stationary sequence data is divided into several samples, and a rolling window with a window size of m is used. The window size refers to the input time period. Each time window constitutes a sample as the input of LSTM, and the sequence value of the next moment of the time window is the target value. After training with the training set, the parameters of LSTM can be obtained, including the weight matrix and the offset term. Therefore, the single-step prediction of the non-stationary subsequence of the gyro random error can be defined as using the non-stationary subsequence values ​​of the current moment and the previous m-1 moments, a total of m data, to predict the non-stationary subsequence value of the next moment, as follows: in, Represents the LSTM model. By training the LSTM network, the dynamic characteristics of the non-stationary subsequence can be learned, and the non-stationary subsequence value at the next moment can be accurately predicted, and the internal state of the network can be continuously updated.

[0056] Step 5: Based on the third and fourth steps, the hybrid model of MEMS gyro random error can be obtained Since the model of random error is a nonlinear function, the cubature Kalman filter CKF is used to realize the real-time estimation and compensation of random error. The specific implementation is as follows:

[0057] According to the third and fourth steps, the hybrid model of MEMS gyroscope random error can be obtained: in, Represents the random error prediction value of the hybrid model at time k; Represents the ARMA model of the ith stationary subsequence; LSTM model representing non-stationary subsequences; in order to compensate for the random error of MEMS gyroscope in real time, the hybrid model is introduced into the filtering framework. Considering that too high a model order will lead to a sharp increase in filtering calculations, the system state is defined according to the network model , is the non-stationary subsequence of the random error of MEMS gyroscope.

[0058] The system state equation and measurement equation can be expressed as: in, Represents the one-step prediction error of the network model. is a nonlinear function, so the state equation is a nonlinear equation; is the observed value of the MEMS gyro under static (zero input) conditions at time k. After compensating the predicted values ​​of all stationary subsequences, the measured value of the non-stationary subsequence containing white noise can be obtained. .

[0059] The system equations can be simplified to: in, ; is the measurement matrix; represents the process noise, The covariance of ; represents a nonlinear transfer function; represents the measurement noise.

[0060] The steps of the MEMS gyroscope multi-source error separation estimation method based on the hybrid model are as follows:

[0061] (1) Prediction stage

[0062] right The optimal estimate of the state at the moment Sampling from the posterior probability distribution: In the formula For the Volume points; and Respectively The posterior state estimate and error covariance matrix at time t; Represents the square root matrix obtained by Cholesky decomposition of the matrix List.

[0063] (2) Time update phase

[0064] The sample points are transformed by nonlinear functions and the predicted mean and covariance are calculated.

[0065] a) The volume points realize state transfer through nonlinear transfer function: The LSTM network in the hybrid model is When transferring volume points, no internal state update is performed. After the state is transferred, the optimal estimate of the state at the previous moment As the input of LSTM, the internal state of LSTM network is updated to be the volume point at the next moment Prepare for the transfer.

[0066] b) Weighted calculation of the prior probability distribution of the state variables at time k: (3) Measurement update phase Therefore, the optimal state estimate at time k is and its error covariance for: in .

[0067] In summary, by establishing a hybrid model for the random error of the MEMS gyroscope and introducing the CKF filtering framework, the optimal prediction of the random error at time k is obtained. .

[0068] In the real-time filtering process, the estimated random error can be used to compensate the raw output of the gyro , the autocorrelation of the residual after compensation is close to the Dirac function, indicating that the useful information in the original sequence has been extracted.

[0069] Embodiment 1:

[0070] Example 1 Taking actual MEMS gyro measurement as an example, the experimental equipment includes a power supply, a MEMS inertial measurement unit ADIS16465, and a host computer. ADIS16465 contains a three-axis digital gyro, and the gyro zero bias stability is , the angle random walk is . First, the IMU is placed statically on the test bench and preheated to a stable temperature. The sampling frequency is 100 Hz, and a total of 3000 points are collected. The data is divided into a training set and a test set. The first 2000 points are used to obtain a one-step prediction model for random errors. The same training set data is used to obtain LSTM network models, ARMA models and hybrid models. The last 1000 points are used as real-time measurement values ​​to perform online estimation of random errors using the method proposed in the present invention. First, the MEMS gyro error data is preprocessed, and the mean filter is used to eliminate the white noise in the sequence data and the window size is set to 0.1s.

[0071] Since the gyro random error data is a discrete signal, it meets the conditions of discrete wavelet transform. The present invention uses Daubechies wavelet 2 as the mother wavelet to compromise between wavelength and smoothness. In order to fully extract the stationary information in the non-stationary sequence, 4 levels of wavelet decomposition are performed. The decomposition diagram is shown in Figure 2 As shown in the figure, 4 detail subsequences and 1 trend subsequence are obtained. KPSS stationarity test is performed on the 5 groups of subsequences respectively, and 4 groups of stationary subsequences and 1 group of non-stationary subsequence are obtained. According to the AIC criterion, the ARMA model orders of the 4 groups of stationary subsequences are determined to be [4,1], [8,1], [6,3] and [10,1] respectively. The LSTM network model is trained using the non-stationary subsequences, and then the hybrid model is obtained by combining them.

[0072] In order to compare the prediction accuracy of the three models, we use the same data set to train and predict the models. The one-step prediction results are as follows: Figure 3 In order to quantitatively compare the prediction accuracy of different models, we selected three statistics for the prediction error, including mean absolute error (MAE), mean square error (MSE) and root mean square error (RMSE), which are specifically defined as: in Represents the one-step prediction value of the model; is the test target value; Represents the amount of data in the test set, here The statistical information of the prediction errors of the three models is shown in Table 1.

[0073] Table 1 Statistics of prediction errors of different models

[0074] method MEAN MSE RMSE Original signal 0.0263 0.0011 0.0337 ARMA model residuals 0.0132 3.2675e-04 0.0181 LSTM model residual 0.0110 2.8730e-04 0.0169 Mixed Model Residuals 0.0089 1.5069e-04 0.0123

[0075] According to the results, the hybrid model has more accurate prediction accuracy than the LSTM network model and the ARMA model. To further verify the performance of the hybrid model-CKF method, we use the test set data without mean filtering as the measurement value, estimate and compensate the random error of the MEMS gyroscope online, compare the proposed method with the ARMA-KF and LSTM-CKF methods, and draw the normalized autocorrelation curves of the signals before and after compensation, as shown in Figure 2. Figure 4 The statistical information of the signal before and after compensation is shown in Table 2.

[0076] Table 2 Statistical information of gyro signals after compensation using different methods

[0077] method MEAN MSE RMSE VAR Original signal 0.0263 0.0011 0.0337 0.0011 ARMA+KF 4.4666e-06 3.5076e-11 5.9225e-06 3.5111e-11 LSTM+CKF 7.4112e-04 1.2320e-06 0.0011 1.2255e-06 Hybrid Model + CKF 3.2473e-04 6.9051e-07 8.3097e-04 6.9027e-07

[0078] As shown in Table 2 and Figure 4 As shown, all three methods can effectively compensate for the gyro random error. By comparing the normalized autocorrelation of the residual, the hybrid model + CKF method can make the compensated residual autocorrelation closer to the Dirac function, that is, closer to white noise, which means that the time-related (useful) information in the original signal is more fully extracted and the gyro error is more fully compensated. The above illustrates the effectiveness of the present invention.

[0079] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, and it should be clear that the present invention is not limited to the scope of the specific embodiments, for those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A method for separating and estimating multi-source errors of a micro-electromechanical gyroscope based on a hybrid model, characterized in that: The following steps are involved: Step 1: Data collection and preprocessing: Collect the output data of the MEMS gyroscope when it is zero input, and preprocess it through mean filtering; Step 2: Error sequence decomposition: Using wavelet transform to decompose the random error sequence of MEMS gyroscope Decompose and extract the trend information and detail information in the sequence; after KPSS stationarity test, the sequence is divided into stationary subsequences and b non-stationary subsequences, where a and b are positive integers greater than or equal to 1; Step 3: Establish an ARMA model for the stationary subsequence: For each stationary subsequence, establish an ARMA random time series model, denoted as ; Step 4: Establish the LSTM network model of non-stationary subsequence: Use the MEMS gyro output data to train the LSTM network model of non-stationary subsequence, denoted as ; Step 5. Error separation estimation based on volumetric Kalman filtering: Combine the two types of models obtained in the third and fourth steps to establish a hybrid model of MEMS gyroscope random errors. , based on the hybrid model, the cubature Kalman filter CKF is used to achieve the real-time separation estimation of gyro random errors.

2. The method for separating and estimating multi-source errors of a MEMS gyroscope based on a hybrid model according to claim 1, characterized in that: In the first step, the MEMS gyro error data is preprocessed before modeling; The constant deviation in the actual acquisition error sequence is compensated, and the mean filtering algorithm is used to eliminate the white noise in the sequence data. The filtering window is set to 0.1s.

3. The method for separating and estimating multi-source errors of a MEMS gyroscope based on a hybrid model according to claim 1, characterized in that: In the second step, in order to model and predict the gyro error sequence, wavelet transform is introduced to extract the stationary components in the non-stationary sequence. The specific steps are as follows: The random error sequence of the MEMS gyroscope after mean filtering is recorded as , After mean filtering The observed value of random error at each moment; The Daubechies wavelet basis function is selected and the discrete wavelet transform is used to decompose the non-stationary sequence. When the first layer of wavelet decomposition is performed, the original signal is passed through a low-pass filter and a high-pass filter respectively to decompose the original signal into a trend sequence. and detail sequence When the next wavelet decomposition is performed, the trend sequence obtained from the previous decomposition is Further decomposed into trend series and detail sequence ; Repeat the above process until the set number of decomposition levels is reached; The wavelet decomposition formula is: in, Indicates After sub-wavelet decomposition, the trend sequence is obtained. is the sequence length; For the The trend sequence after decomposition is is the sequence length; For the The detailed sequence after decomposition, is the sequence length; represents a low-pass filter; represents a high-pass filter; Signal reconstruction is performed for each subsequence so that the subsequence length is consistent with the original sequence. The reconstruction formula is as follows: Perform KPSS stationarity test on each subsequence obtained by decomposition, and get stationary subsequences and non-stationary subsequences, and the sum of the two is the total number of subsequences.

4. The method for separating and estimating multi-source errors of a MEMS gyroscope based on a hybrid model according to claim 3 is characterized in that: b non-stationary subsequences are combined into one non-stationary subsequence by addition, and b is 1 at this time.

5. The method for separating and estimating multi-source errors of a MEMS gyroscope based on a hybrid model according to claim 1, characterized in that: The third step is as follows: The model is represented as follows: in , are the autoregressive order and moving average order of the ARMA model respectively; Represents the predicted value of the stationary subsequence at time t; is the historical observation value of the stationary subsequence; is the historical random interference sequence; and is the model coefficient; is zero-mean white noise; the specific establishment process of the ARMA model is as follows: First, the Akaike information criterion AIC was used to Determine the order of the model by plotting the heat map corresponding to the AIC value to intuitively find the order with the smallest AIC value and , considering the complexity of the model, design and is an integer in the interval [0, 10], and the AIC criterion is expressed as: in is the maximum likelihood estimate of the model residual variance; is the number of observations; After determining the model order, the maximum likelihood estimation method is used to fit the model parameters to obtain the ARMA model.

6. The method for separating and estimating multi-source errors of a MEMS gyroscope based on a hybrid model according to claim 1, characterized in that: The fourth step is as follows: In the LSTM calculation process, each LSTM unit has three gates: input gate, forget gate and output gate; The forget gate is used to control the state of the previous moment to the current memory unit state value Impact: In the formula is the sigmoid activation function; is the input sequence value at time t; and is the weight matrix; is the offset term; is the output of the previous moment; The input gate is used to control the input of the current time step The state value of the memory unit Impact: In the formula , , and is the weight matrix; and is the offset term; tanh is the activation function; For the current moment Bring new information; To combine the current and last updated memory unit states; The output gate determines the output of the current time step: The output value at time t is After scaling with the tanh function The output at this moment is ; Non-stationary subsequences for random errors of MEMS gyroscopes The LSTM model is established. The specific implementation steps are as follows: the collected non-stationary sequence data is divided into several samples, and a rolling window with a window size of m is used. The window size refers to the input time period; each time window constitutes a sample as the input of LSTM, and the sequence value at the moment after the time window is the target value. After training with the training set, the parameters of LSTM can be obtained, including the weight matrix and the offset term; therefore, the single-step prediction of the non-stationary subsequence of the gyro random error can be defined as using the non-stationary subsequence values ​​of the current moment and the previous m-1 moments, a total of m data, to predict the non-stationary subsequence value at the next moment, as follows: in Represents the LSTM model; by training the LSTM network, the dynamic characteristics of the non-stationary subsequence can be learned, and then the non-stationary subsequence value at the next moment can be accurately predicted, and the internal state of the network can be continuously updated.

7. The method for separating and estimating multi-source errors of a MEMS gyroscope based on a hybrid model according to claim 1, characterized in that: The fifth step is as follows: According to the third and fourth steps, the hybrid model of MEMS gyro random error is obtained: in Represents the random error prediction value of the hybrid model at time k; Represents the ARMA model of the ith stationary subsequence; LSTM model representing non-stationary subsequences; In order to compensate for the random errors of MEMS gyroscopes in real time, a hybrid model is introduced into the filtering framework; Considering that a high model order will lead to a sharp increase in filtering calculations, the system state is defined according to the network model. , is the non-stationary subsequence of the random error of the MEMS gyro; The system state equation and measurement equation are expressed as: in, Represents the one-step prediction error of the network model; Represents the observed value of random error at time k after mean filtering, and the measured value of non-stationary subsequence containing white noise is obtained after compensating the predicted value of all stationary subsequences ; The system equations are simplified to: in ; is the measurement matrix; represents the process noise, The covariance of ; represents a nonlinear transfer function; represents the measurement noise; The steps of the MEMS gyro random error separation estimation and compensation algorithm based on the hybrid model are as follows: (1) Prediction stage right The optimal estimate of the state at the moment Sampling from the posterior probability distribution: In the formula For the Volume points; and Respectively The posterior state estimate and error covariance matrix at time t; represents the square root matrix obtained by Cholesky decomposition of the matrix List; (2) Time update phase Transform the sample points through nonlinear functions and calculate the predicted mean and covariance; a) The volume points realize state transfer through nonlinear transfer function: The LSTM network in the hybrid model is When transferring volume points, the internal state is not updated; when completing the current moment After the state is transferred, the optimal estimate of the state at the previous moment As the input of LSTM, the internal state of LSTM network is updated to be the volume point at the next moment Prepare for the transfer; b) Weighted calculation of the prior probability distribution of the state variables at time k: (3) Measurement update phase Therefore, the optimal state estimate at time k is and its error covariance for: in ; In summary, by establishing a hybrid model for the random error of the MEMS gyroscope and introducing the CKF filtering framework, the optimal prediction of the random error at time k is obtained. ; The estimated random error is used to compensate the raw output of the gyro during real-time filtering. , the autocorrelation of the residual after compensation is close to the Dirac function.