An online estimation algorithm for random errors of vehicle-mounted inertial devices

Abstract

The present invention discloses an online estimation algorithm for random errors of vehicle-mounted inertial devices. The algorithm performs parking detection based on the output data of the vehicle-mounted inertial sensor. When the vehicle is parked, static inertial data is accumulated, and after removing the IMU zero bias, generalized wavelet moment estimation is used to perform online modeling and parameter estimation of random errors. When driving, the algorithm uses an INS / GNSS integrated navigation system for navigation and positioning, estimates the IMU zero bias, and uses a random noise model constructed during parking to compensate for the inertial output data. The compensation result is fed into the navigation system for solution. As the number of parking times increases, the accuracy of the random noise model constructed by the algorithm increases accordingly. This invention overcomes the limitations of existing inertial sensor calibration methods, implements online modeling of random noise, and improves the positioning accuracy of the INS / GNSS integrated navigation system.

Description

Technical Field

[0001] The present invention relates to the technical field of inertial device calibration, and more particularly to an online modeling and parameter estimation algorithm for random errors of inertial devices. Background Art

[0002] Accurately, continuously, and reliably estimating the position, velocity, and attitude of moving objects in space, such as vehicles, aircraft, and ships, is a critical issue for practical applications such as autonomous driving, drone operations, and precision agriculture. Currently, navigation and positioning using the Global Navigation Satellite System (GNSS) is the most commonly used solution in urban environments. However, when moving objects are in complex urban environments, the partial or complete unavailability of satellite signals due to overpasses, tunnels, or tree shade can severely degrade the performance of such systems. Furthermore, for certain specific needs, the bandwidth of GNSS satellite receivers is relatively low (typically less than 10Hz), making it impossible to fully rely on the navigation parameters provided by GNSS. Furthermore, GNSS does not provide information about attitude.

[0003] To address these challenges, a common and widely used approach is to combine GNSS with an inertial navigation system (INS). Due to their complementary strengths, combining GNSS and INS can significantly improve navigation and positioning performance. When GNSS signals are available, the integration of INS and GNSS is typically achieved through Bayesian techniques, with commonly used algorithms including standard Kalman filtering, variations of Kalman filtering, and robust adaptive filtering. When GNSS signal quality is poor or completely unavailable, the INS operates in a recursive mode, meaning that the navigation state can be estimated completely independently of GNSS. In both modes, overall navigation performance depends heavily on the accuracy of the inertial signal, or more precisely, on its errors. These errors are integrated into the INS for computation, and their impact increases dramatically over time. Therefore, accurately modeling and estimating inertial signal errors is crucial for correctly estimating and improving navigation performance.

[0004] Currently, traditional methods for modeling inertial device error signals, such as the Allan variance method and PSD analysis, suffer from the inability to identify and separate errors in the spectral domain. The specific analysis is as follows:

[0005] (1) The AV method is only applicable to noise processes that can be clearly identified and separated in the spectral domain and are not affected by spectral ambiguity. Moreover, the AV method does not allow direct reading of the parameters of the GM process, because larger values ​​make the process similar to WN, while smaller values ​​make the process similar to RW. Therefore, the traditional AV method is limited to models consisting of processes characterized by linear regions in the wavelet variance logarithm plot. In addition, in most cases, due to the interference of many influencing factors during the sensor data acquisition process, its wavelet variance logarithm plot cannot show the characteristics of the typical linear region. Therefore, in most cases, it is difficult to use the AV method for coefficient reading in practice.

[0006] (2) The calculation of empirical WV is simpler than parameter-free PSD analysis. For example, the periodogram is an inconsistent estimator of the power spectral density function and can be severely biased even for large sample sizes due to frequency leakage effects. Therefore, more complex PSD estimators or smoothing techniques such as pre-whitening or tapering are needed to approach the consistency provided by generalized wavelet moment estimates;

[0007] (3) When the PSD has large variability over a very narrow frequency band, it will make the optimization problem based on the difference between the empirical PSD and the model-based PSD more difficult to solve.

[0008] Therefore, based on the above background, how to overcome the limitations of traditional methods and provide an online estimation algorithm for random errors of inertial devices is an urgent problem that needs to be solved by researchers in the field of navigation and positioning. Summary of the Invention

[0009] In view of this, the present invention provides an online estimation algorithm for the random error of a vehicle-mounted inertial device, which is used to solve the problem of online modeling and parameter estimation of the random error of an inertial device of a specific vehicle-mounted object during driving.

[0010] In order to achieve the above object, the present invention adopts the following technical solutions:

[0011] An online estimation algorithm for random errors of vehicle-mounted inertial devices includes the following steps:

[0012] S1. Determine whether the vehicle is in a parked state based on the IMU observation data output by the vehicle-mounted object.

[0013] If not, use the INS / GNSS integrated navigation system for navigation and positioning, estimate the IMU accelerometer bias, compensate for random noise, and send the compensation results to the navigation system for solution;

[0014] If yes, proceed to step S2;

[0015] S2. Obtain parking section IMU observation data and accelerometer zero bias, accumulate the parking section IMU observation data, subtract the accelerometer zero bias, and obtain a random error component;

[0016] S3. Estimating parameters of the random error component using generalized wavelet moment estimation, compensating for random noise based on the parameter estimation, and sending the compensation result to the navigation system for solution;

[0017] Preferably, a parking detection criterion is established based on the static output data of the IMU.

[0018] The parking inspection criteria are:

[0019]

[0020] Where A i is the data output by the accelerometer at time i, U i is the mean of the data in the fixed time window at time i, the number of data in the fixed time window is N, T i is the standard deviation of the data at time i, and λ is the test threshold;

[0021] Preferably, the test threshold λ is dynamically adjusted by the following formula:

[0022]

[0023] In the formula, k is the number of times the parking point is detected, and the initial threshold is selected as T based on experience. i =0.01;

[0024] Preferably, the random error component is formed by a linear combination of independent random processes, and the independent random processes include: Gaussian white noise (WN), random walk (RW), random ramp (RR), quantization noise (QN) and first-order autoregressive process (AR);

[0025] Preferably, before executing step S3, v(τ j )~τ j Performing log-log processing on the curve to obtain a double logarithmic curve, constructing a plurality of candidate models of the random error component according to the slope characteristics of the double logarithmic curve, and performing parameter estimation on the plurality of candidate models of the random error component using generalized wavelet moment estimation;

[0026] Preferably, a selection criterion is constructed to determine the optimal parameter estimates of the plurality of candidate models of the random error component, wherein the selection criterion is:

[0027]

[0028] In the formula, θ is the parameter to be estimated, is the wavelet variance obtained from the observation sequence, is the wavelet variance calculated according to the model, Ω is the positive definite weight matrix that makes the formula convex;

[0029] The model with the smallest GOF value is selected as the optimal model, and the parameter estimate corresponding to the optimal model is the optimal parameter estimate;

[0030] Preferably, for the first segment of parking data, multiple candidate models of the random error components are constructed, and for the nth segment of parking data, where n≥1, the optimal model is directly used for parameter estimation;

[0031] Preferably, the expression for performing parameter estimation on the random error component using generalized wavelet moment estimation is:

[0032]

[0033] In the formula, θ is the parameter to be estimated, is the wavelet variance obtained from the observation sequence, φ(θ) is the wavelet variance calculated according to the model, and Ω is the positive definite weight matrix that makes the formula convex;

[0034] Where Ω is a positive definite weight matrix that makes the equation convex.

[0035] As can be seen from the above technical solution, compared with the prior art, the present invention discloses an online estimation algorithm for the random error of a vehicle-mounted inertial device. This application uses wavelet variance and least squares method to perform online parameter estimation of the random error of the inertial device to improve the real-time navigation and positioning accuracy of the vehicle on the object;

[0036] Another object of the present invention is to construct a random error candidate model and a model selection criterion to achieve automatic modeling and parameter estimation of random noise, so as to further improve the accuracy of parameter estimation;

[0037] Another object of the present invention is to accumulate static data of parking status when obtaining random error components. As the number of parking times increases, the amount of IMU static data increases, and the modeling of IMU random noise becomes more and more accurate.

[0038] Another object of the present invention is to construct and traverse a candidate model only for the first segment of parking data. When static data accumulation is performed when the parking state is detected, the candidate model with the optimal random error component can be directly used for parameter estimation, thereby improving estimation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0040] Figure 1 The attached figure is a flow chart of the online estimation algorithm for random errors of vehicle-mounted inertial devices.

[0041] Figure 2 The attached figure shows the v(τ j )~τ j The slope characteristic graph presented by the double logarithmic curve graph;

[0042] Figure 3 The accompanying figure is a flow chart of the algorithm for automatically selecting the optimal model based on GMWM. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] Since the random error components of low-cost inertial sensors are quite complex, commonly used random error analysis methods such as the Allan variance method and PSD analysis are unable to identify and separate errors in the spectral domain. To this end, the present invention uses generalized wavelet moment estimation to estimate random errors, overcoming the limitations of traditional methods. Static data is accumulated through parking state detection of the vehicle carrier, and online modeling and parameter estimation of the IMU's random errors are performed when the vehicle is stationary, thereby improving the positioning accuracy of the integrated navigation system including INS.

[0045] The generalized method of wavelet moments (GMWM) links the wavelet variance with the PSD and represents the sequence to be estimated as a model consisting of the sum of Gaussian white noise (WN), random walk (RW), random ramp (RR), quantization noise (QN), and a finite number of first-order autoregressive processes (AR). The GMWM estimator constructed under this model is asymptotically consistent. The wavelet variance of the sequence to be estimated is then matched to the wavelet variance implied by the assumed model, and the generalized least squares method is used to minimize the difference between the two to estimate the parameters of the latter. This method can effectively avoid the drawbacks of the above traditional methods and has strong practical application value.

[0046] Specifically, the embodiment of the present invention discloses an online estimation algorithm for random errors of vehicle-mounted inertial devices, such as Figure 1 As shown, the following steps are included:

[0047] S1. Determine whether the vehicle is in a parked state based on the IMU observation data output by the vehicle-mounted object.

[0048] If not, use the INS / GNSS integrated navigation system for navigation and positioning, estimate the IMU accelerometer bias, compensate for random noise, and send the compensation results to the navigation system for solution;

[0049] If yes, proceed to step S2;

[0050] S2. Obtain parking section IMU observation data and accelerometer zero bias, accumulate the parking section IMU observation data, and subtract the accelerometer zero bias to obtain a random error component;

[0051] S3. Parameter estimation is performed on the random error component using generalized wavelet moment estimation, random noise is compensated according to the parameter estimation, and the compensation result is sent to the navigation system for solution.

[0052] For in-vehicle navigation systems, the vehicle itself contains specific constraints in different states of motion. In particular, when the vehicle is parked, its ground velocity is zero and its attitude remains unchanged. In this case, zero velocity and zero angular rate can be used as constraints. This information provides additional observations for the navigation system, helping to improve the accuracy and stability of the integrated navigation system. It is particularly important for suppressing accumulated inertial navigation errors when GNSS signals are unavailable.

[0053] In this invention, we first establish parking detection criteria based on the potential relationship between the output of the inertial device under different motion states and the motion state, taking advantage of the statistical characteristics of the accelerometer data of the IMU (Inertial Measurement Unit) in the stationary state. Specifically, when the vehicle is parked, because the inertial device output data is not affected by vehicle maneuvers, its output data is relatively stable. Therefore, stability analysis of the acceleration output data sequence can be performed to determine the parking state. A common method is to use the standard deviation of the accelerometer data within a fixed time window as the test statistic. The judgment criterion is:

[0054]

[0055] Where A i is the data output by the accelerometer at time i, U i is the mean of the data in the fixed time window at time i, the number of data in the fixed time window is N, T i is the standard deviation of the data at time i, and λ is the test threshold;

[0056] In one embodiment, 1 s is selected as the time window length, and the accelerometer sampling frequency is 100 Hz, then N=100.

[0057] Since different types of inertial sensors have different specifications and their output data standard deviations in a static state are also different, the use of empirical thresholds is prone to misjudgment. Moreover, for the situation studied in this invention, it is necessary to ensure that the IMU data used for GMWM estimation modeling is in a static state. False detection has a greater impact on the accuracy of modeling estimation than missed detection. Therefore, the test threshold λ in this invention is dynamically adjusted by the following formula:

[0058]

[0059] In the formula, k is the number of times the parking point is detected, and the initial threshold is selected as T based on experience. i =0.01.

[0060] The physical meaning of the first item is: when the data standard deviation is less than the test threshold again, it is more likely to be considered that the vehicle is in a continuously parked state, so the detection threshold is increased, making it easier to detect the parked state;

[0061] The physical meaning of the second item is: if the standard deviation of the data at the current moment is too different from that at the previous moment, it indicates that the vehicle's motion state has changed, and it is more inclined to believe that the vehicle is not in a parked state. Therefore, the detection threshold is reduced, making it more difficult to detect the parked state.

[0062] Use the above criteria to determine whether the vehicle-borne object is in a parked state.

[0063] If the vehicle is not in a parked state, that is, the vehicle is moving, the INS / GNSS integrated navigation system is used to provide navigation and positioning services and estimate the accelerometer bias of the IMU;

[0064] If the vehicle-borne object is in a parking state, obtain the parking section IMU observation data and the accelerometer zero bias, accumulate the parking section IMU observation data, and subtract the accelerometer zero bias to obtain the random error component.

[0065] The IMU static data obtained during this parking period is accumulated with all previous static data. As the number of parking times increases, the amount of IMU static data will increase, and GMWM's modeling of IMU random noise will become more and more accurate.

[0066] In one embodiment, before performing parameter estimation on the random error component using generalized wavelet moment estimation, a candidate model of the random error component is first constructed.

[0067] Specifically, the random error component is composed of a linear combination of independent random processes, including Gaussian white noise (WN), random walk (RW), random ramp (RR), quantization noise (QN), first-order autoregressive process (AR) and flicker noise, and the GMWM estimation based on these basic random processes satisfies a consistent and asymptotically normal distribution.

[0068] When using Haar wavelet filter, the analytical expressions of WN, QN, RW, RR and multiple AR processes are shown in Table 1.

[0069] Table 1 PSD and discrete difference state equation of the basic random process

[0070]

[0071] According to the relationship between variance and power spectral density function Take the logarithm with base 2 on both ends of the formula, and j )~τ j The typical random processes listed in Table 1 are plotted in a double logarithmic curve, and it can be seen that they exhibit special slope characteristics such as Figure 2 The slope characteristics are shown in Table 2.

[0072] The above process can be referred to IMU Noise Parameter Identification - Allan Variance (https: / / zhuanlan.zhihu.com / p / 158927004)

[0073] Table 2 Slopes of WV of the underlying random process in the double logarithmic plot

[0074] Model QN WN AR RW Drift(RR) WV -2 -1 -1～1 1 2

[0075] Furthermore, according to the slope characteristics of the double logarithmic curve, multiple random error component candidate models are constructed. k :k=1,…,N}, according to its v(τ j )~τ j The slope characteristics of the double logarithmic curve are used to infer its possible composition, which is constructed as a linear combination of QN, WN, RW, RR, and k AR models. This candidate model can be inferred manually or the optimal model can be automatically selected by constructing a random error population model and a preferential ranking criterion. For inertial devices, the most commonly used model to describe its random noise is:

[0076] x t+1 =e -βΔt x t +w t+1 +u t+1

[0077] y t+1 =x t+1 +v t+1

[0078] Where Δt is the time interval of observation, u t =ωΔt,w t ~N(0,q) and The sequence is composed of a first-order Gauss-Markov model, a drift model and a white noise model, that is, error = GM + DR + WN.

[0079] In one implementation, in order to cover all possible models in the widest range, the present invention defines the IMU random noise overall model as:

[0080] error=5*AR+DR+WN+QN+RW

[0081] Among them, Random Ramp (RR) is also called Drift (DR).

[0082] When training, all models formed by combining all the basic random processes in the model are considered as candidate models.

[0083] Then, generalized wavelet moment estimation is used to perform parameter estimation on the plurality of candidate models of the random error component.

[0084] Since the IMU observation data used for online estimation is accumulated during the parking period and the data volume is small, parameter estimation for all candidate models is quite fast, and the calculation time can generally be completed within a few seconds.

[0085] Furthermore, the wavelet variance (WV) can be understood as the variance of a random process after passing through an approximate bandpass filter, which is calculated by the maximum overlap discrete wavelet transform (MODWT), and its wavelet coefficients are filtered using wavelet Construct, for j = 1, MODWT satisfies,

[0086]

[0087] In the formula, for l<0,l≥L1, L1 is The length of b, m is a non-zero integer. The transfer function is:

[0088]

[0089] The j-th wavelet filter {h j,l The length L of j =(2 j -1)(L1-1)+1 can be obtained by calculating the inverse Fourier transform of the formula,

[0090]

[0091] MODWT is actually a discrete wavelet transform filter h j,l A scaled version of For a finite sequence {y k :k=1,…,N}Use The MODWT wavelet coefficients can be obtained:

[0092]

[0093] Define the wavelet variance as τ j =2 j-1 Sequence on scale Variance of:

[0094]

[0095] Wavelet coefficient sequence is stationary, and its power spectral density (PSD) is:

[0096]

[0097] In the formula, |·| represents the modular operation, F θ The model expression F(θ) for generating data is:

[0098]

[0099] Formula (7) shows that the variance of the wavelet coefficient sequence is equal to the integral of its power spectrum density, that is, there is an implicit relationship between the wavelet variance and the parameter θ of the model F(θ). Therefore, this implicit relationship can be explored by defining a suitable estimator of θ, that is, by matching v(τ j ) and the model-based v(τ j ) expression to estimate the parameter θ. , so there exists a mapping:

[0100]

[0101] For a finite sequence {y k :k=1,…,N}, the wavelet variance estimator of MODWT is

[0102]

[0103] Where, And M j =NL j +1. The estimator is v(τ j ) consensus estimate.

[0104] The core idea of ​​GMWM is to find the implicit parameter estimates from the empirical WV obtained from the observation sequence, that is,

[0105]

[0106] Where θ(v) is the inverse function of the theoretical wavelet variance v(θ), is the wavelet variance obtained from the observation series using the MODWT estimator.

[0107] In most cases, the mapping shown in Equation (8) is implicit, making it difficult to find the inverse function θ(v). However, the difference between the empirical WV obtained from the observed sequence and the theoretical WV contained in the model F(θ) should be minimized, a concept that coincides with the least squares principle. The essence of the least squares estimation method is to establish an estimate such that the difference between the actual value and the estimated value is as small as possible, and to measure this difference using the square of the difference between the actual value and the estimated value.

[0108] Therefore, the wavelet variance and the generalized least squares principle are combined to construct the estimator. The WV obtained from the model is recorded as φ(θ), and the WV obtained from the observation sequence is recorded as The GLS optimization problem is calculated as shown in formula (11):

[0109]

[0110] In the formula, θ is the parameter to be estimated, is the wavelet variance obtained from the observation sequence, φ(θ) is the wavelet variance calculated according to the model, Ω is the positive definite weight matrix that makes the formula convex, and formula (11) is the constructed GMWM estimator.

[0111] As mentioned above, for WN, QN, RW, RR and multiple AR models and their combination processes, they are consistent and therefore obey the distribution.

[0112]

[0113] Where, And B=(D T ΩD) -1 D T Ω, matrix When Ω=I,

[0114]

[0115] When choosing The most concise estimator expression can be obtained:

[0116]

[0117] When using the Haar wavelet filter, the WV of the composite process formed by the linear combination of multiple independent processes is equal to the sum of the WVs of the independent processes that make up the composite process, that is, formula (7) can be expanded to

[0118]

[0119] Where m is the number of independent processes, It is PSD, v m (τ j ) is in the process {(Y m ) k} scale τ j When the analytical expression of v(θ) is known, the parameter estimation can be performed using formula (11).

[0120] The generalized wavelet moment estimation is used to estimate the parameters of the candidate models of the random error components, and multiple parameter estimation results are obtained. At this time, it is necessary to optimize the multiple parameter estimation results. Correspondingly, the objective function given by formula (11) is to measure the φ(θ) obtained by the model calculation and the observed data. The measure of the difference between them, minimizing this difference makes the model closer to the observation sequence. Based on this physical meaning, the Goodness of Fitness (GOF) is defined as:

[0121]

[0122] In the formula, θ is the parameter to be estimated, is the wavelet variance obtained from the observation sequence, is the wavelet variance calculated according to the model, Ω is the positive definite weight matrix that makes the formula convex,

[0123] After the parameters of all candidate models are estimated using GMWM, all candidate models are sorted based on the GOF criterion, and the model with the smallest GOF value is selected as the optimal model. The flowchart of the entire automatic selection of the optimal model is as follows Figure 3 shown.

[0124] In one embodiment, a plurality of candidate models of the random error components are constructed only for the first segment of parking data, and the optimal model is directly used to perform parameter estimation for the nth segment of parking data, where n≥1.

[0125] Because for the same inertial device, the characteristics of its static random noise are stable. As time goes by, the model structure of the random noise remains unchanged, and only its parameters change with time. Therefore, this operation of traversing all candidate models only needs to be performed when obtaining the first segment of parking data. When the parking state is detected later and static data is accumulated, there is no need to perform finalization, and GMWM can be used directly for parameter estimation.

[0126] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0127] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An online estimation algorithm for random errors of vehicle-mounted inertial devices, characterized in that: The steps include: S1. According to the IMU observation data output by the vehicle-mounted object, determine whether the vehicle is in a parked state. The test threshold λ is dynamically adjusted according to the following formula: In the formula, k is the number of times the parking point is detected, and the initial threshold is selected as T based on experience. i =0.01; If not, use the INS / GNSS integrated navigation system for navigation and positioning, estimate the IMU accelerometer bias, compensate for random noise, and send the compensation results to the navigation system for solution; If yes, execute steps S2-S3; S2. Obtain parking section IMU observation data and accelerometer zero bias, accumulate the parking section IMU observation data, and subtract the accelerometer zero bias to obtain a random error component; the random error component is formed by a linear combination of independent random processes; S3, using generalized wavelet moment estimation to perform parameter estimation on the random error component, including constructing a selection criterion to determine optimal parameter estimates of multiple candidate models of the random error component, The selection criteria are: In the formula, θ is the parameter to be estimated, is the wavelet variance obtained from the observation sequence, is the wavelet variance calculated according to the model, Ω is the positive definite weight matrix that makes the formula convex; Random noise is compensated according to the parameter estimation, and the compensation result is sent to the navigation system for solution.

2. The online estimation algorithm for random errors of vehicle-mounted inertial devices according to claim 1, characterized in that: According to the IMU observation data, parking inspection criteria are established. The parking inspection criteria are: Where A i is the data output by the accelerometer at time i, U i is the mean of the data in the fixed time window at time i, the number of data in the fixed time window is N, T i is the standard deviation of the data at time i, and λ is the test threshold.

3. The online estimation algorithm for random errors of vehicle-mounted inertial devices according to claim 1, characterized in that: The independent random process includes: Gaussian white noise WN, random walk RW, random ramp RR, quantization noise QN and first-order autoregressive process AR.

4. The online estimation algorithm for random errors of vehicle-mounted inertial devices according to claim 3 is characterized in that: Before executing step S3, for each of the independent random processes v(τ j )~τ j The curve is processed by log-logarithm to obtain a double logarithmic curve. According to the slope characteristics of the double logarithmic curve, a plurality of candidate models of the random error component are constructed, and parameters of the plurality of candidate models of the random error component are estimated using generalized wavelet moment estimation.

5. The online estimation algorithm for random errors of vehicle-mounted inertial devices according to claim 4 is characterized in that: The model with the smallest GOF value is selected as the optimal model, and the parameter estimate corresponding to the optimal model is the optimal parameter estimate.

6. The online estimation algorithm for random errors of vehicle-mounted inertial devices according to claim 5, characterized in that: For the first section of parking data, multiple candidate models of the random error components are constructed, and for the nth section of parking data, n≥1, the optimal model is directly used to perform parameter estimation.

7. The online estimation algorithm for random errors of vehicle-mounted inertial devices according to claim 1, characterized in that: The expression for parameter estimation of the random error component using generalized wavelet moment estimation is: In the formula, θ is the parameter to be estimated, is the wavelet variance obtained from the observation sequence, φ(θ) is the wavelet variance calculated according to the model, and Ω is the positive definite weight matrix that makes the formula convex.

Images