CN117609737B - Method, system, equipment and medium for predicting health state of inertial navigation system - Google Patents

Abstract

The invention discloses a method, system, equipment and medium for predicting the health state of an inertial navigation system, and relates to the field of health management of the inertial navigation system. The method includes: obtaining test data of the current period of the inertial navigation system; the test data is non-periodic data. ; Use the Markov Monte Carlo method to interpolate the missing values in the test data of the current period to obtain the periodic data of the current period; perform dimensionality reduction on the periodic data of the current period to obtain the periodic dimensionality reduction data of the current period; Input the periodic dimensionality reduction data of the current period into the health state prediction model to obtain the health state of the inertial navigation system in the current period; the health state prediction model is constructed based on machine learning methods. The invention can improve the accuracy of health state prediction of the inertial navigation system.

Description

A method, system, device and medium for predicting the health status of an inertial navigation system

Technical Field

The present invention relates to the field of inertial navigation system health management, and in particular to an inertial navigation system health state prediction method, system, equipment and medium.

Background Art

As a key component of the control system, the inertial navigation system plays a role in precise positioning and attitude determination in complex dynamic systems. It is one of the high-precision components of the system. It plays a vital role in the fields of space shuttles and launch vehicles. The purpose of predicting the health status of the inertial navigation system is to comprehensively use historical information and test data to evaluate the overall performance and status of the system. This evaluation is of great significance and can effectively evaluate the performance and status of the system, identify potential risks of the system, and achieve fault diagnosis and preventive maintenance at a low cost.

However, in the actual application of inertial navigation systems, due to the limitation of the number of tests, there is a lack of high-value health status samples that can be obtained by inertial navigation systems. At the same time, due to the non-periodic test time interval, it is impossible to obtain periodic and continuous health status data of the inertial navigation system. In the case of small data volume and lack of continuous detection data, evaluation errors will continue to accumulate over time, which may lead to a decrease in the accuracy of the assessment of the health status and potential problems of the equipment or even insufficient assessment. The existence of the above problems may have a negative impact on the reliability, performance and maintenance strategy of the equipment.

When predicting the health status of an inertial navigation system, data detected at equal intervals has better continuity and stability in sampling frequency. This continuity and stability helps reduce data noise and uncertainty, improves the stability and accuracy of predictions, and thus builds a more accurate model and improves prediction results. Therefore, in order to more accurately evaluate the health status of the equipment, it is necessary to consider periodic processing of non-periodic test data to compensate for the limitations of limited test data and time interval uncertainty.

Summary of the invention

Based on this, the embodiments of the present invention provide a method, system, device and medium for predicting the health status of an inertial navigation system to solve the limitations caused by limited test data and uncertainty in time intervals, thereby improving the accuracy of predicting the health status of an inertial navigation system.

To achieve the above objectives, the embodiments of the present invention provide the following solutions.

A method for predicting the health status of an inertial navigation system comprises: obtaining test data of a current period of the inertial navigation system; the test data being non-periodic data; interpolating missing values in the test data of the current period using a Markov Monte Carlo method to obtain periodic data of the current period; performing dimension reduction on the periodic data of the current period to obtain periodic dimension reduction data of the current period; inputting the periodic dimension reduction data of the current period into a health status prediction model to obtain the health status of the inertial navigation system in the current period; wherein the health status prediction model is constructed based on a machine learning method.

Optionally, the method for determining the health status prediction model includes: obtaining test data and corresponding health status of the inertial navigation system for a historical period; using the Markov Monte Carlo method to interpolate missing values in the test data for the historical period to obtain periodic data for the historical period; reducing the dimension of the periodic data for the historical period to obtain periodic reduced-dimensional data for the historical period; constructing training data based on the periodic reduced-dimensional data for the historical period and the corresponding health status; using the training data to train a support vector machine, and determining the trained support vector machine as the health status prediction model.

Optionally, the Markov Monte Carlo method is used to interpolate missing values in the test data of the current period to obtain the periodic data of the current period, specifically including: using the Metropolis-Hastings algorithm and the method of estimating potential scale reduction factors to interpolate missing values for any feature in the test data of the current period to obtain the periodic data of each feature of the current period; a parameter in the test data is used as a feature; and the periodic data of all features of the current period are determined as the final periodic data of the current period.

Optionally, the Metropolis-Hastings algorithm and the method for estimating potential scale reduction factors are used to interpolate missing values for any feature in the test data of the current period to obtain periodic data of each feature in the current period, specifically including: for any feature in the test data of the current period, the process of interpolating missing values includes: constructing an initial Markov chain of the feature that obeys a stationary distribution according to a time series; determining the acceptance probability of state transition according to a proposed distribution and a stationary distribution; generating random samples of missing positions according to the acceptance probability; inserting the random samples into the missing positions of the initial Markov chain to generate a new Markov chain; using the method for estimating potential scale reduction factors to determine whether the new Markov chain converges; if converged, inserting the random samples into the test data of the current period to obtain periodic data of the feature in the current period; if not converged, reconstructing the initial Markov chain of the feature that obeys a stationary distribution until a converged new Markov chain is generated.

Optionally, the periodic data of the current period is reduced in dimension to obtain the periodic reduced-dimension data of the current period, specifically including: using a principal component analysis method to reduce the dimension of the periodic data of the current period to obtain the periodic reduced-dimension data of the current period.

Optionally, the test data includes: four types of parameter sets, namely, the zero-order drift coefficient set of the accelerometer, the first-order drift coefficient set of the accelerometer, the zero-order drift coefficient set of the gyroscope and the first-order drift coefficient set of the gyroscope; each type of parameter set includes corresponding drift coefficients on different directional axes; the drift coefficient corresponding to one directional axis is taken as a parameter.

The present invention also provides an inertial navigation system health status prediction system, including: a data acquisition module, used to obtain test data of the inertial navigation system in the current time period; the test data is non-periodic data; an interpolation module, used to interpolate missing values in the test data of the current time period using the Markov Monte Carlo method to obtain periodic data of the current time period; a dimension reduction module, used to reduce the dimension of the periodic data of the current time period to obtain periodic reduced dimension data of the current time period; a health status prediction module, used to input the periodic reduced dimension data of the current time period into a health status prediction model to obtain the health status of the inertial navigation system in the current time period; wherein the health status prediction model is constructed based on a machine learning method.

The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the above-mentioned inertial navigation system health status prediction method.

The present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program implements the above-mentioned inertial navigation system health status prediction method when executed by a processor.

According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the embodiments of the present invention adopt the Markov Monte Carlo method to convert the non-periodic test data in the inertial navigation system into periodic data by interpolating missing values, and based on the periodic data, a health status prediction model is constructed in combination with a data dimension reduction method and a machine learning-based method to realize the prediction of the health status of the inertial navigation system, which solves the limitations caused by limited test data and uncertainty in time intervals, and can improve the accuracy of the health status prediction of the inertial navigation system, thereby positively affecting the reliability, performance and maintenance strategy of the equipment.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

FIG1 is a flow chart of a method for predicting the health status of an inertial navigation system provided by an embodiment of the present invention;

FIG2 is a comparison diagram of test data of the zero-order drift coefficient of the X-axis of the gyroscope provided by an embodiment of the present invention before and after periodization;

FIG3 is a schematic diagram of variance ratio after PCA dimensionality reduction provided by an embodiment of the present invention;

FIG. 4 is a structural diagram of an inertial navigation system health status prediction system provided by an embodiment of the present invention.

DETAILED DESCRIPTION

The technical solutions in the embodiments of the present invention will be described below in conjunction with the accompanying drawings in the embodiments of the present invention.

In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

Embodiment 1

This embodiment provides a method for predicting the health status of an inertial navigation system. For non-periodic test data that can reflect the health status of the inertial navigation system, the Metropolis-Hastings algorithm (M-H) algorithm is used to extract a random sample sequence from a probability distribution when it is difficult to directly sample, and the missing values are interpolated with the most likely values, so as to obtain periodic data while retaining the original data. The M-H algorithm is a Markov Monte Carlo method in statistical physics. Secondly, in order to enhance the accuracy and reliability of the prediction results, the principal component analysis method (Principal Components Analysis, PCA) is used to reduce the dimension and reduce the redundancy and noise in the data. Finally, the support vector machine (SVM) is used to learn and construct a prediction model based on the periodic data after dimension reduction, so as to predict the health status of the inertial navigation system at future moments.

The following is a detailed description of the inertial navigation system health status prediction method of this embodiment.

1 , the inertial navigation system health status prediction method of this embodiment includes the following steps.

Step 101: Acquire test data of the inertial navigation system in the current period; the test data is non-periodic data.

As important components of the inertial navigation system, the accelerometer and gyroscope are installed in the XYZ axis direction of the inertial navigation system. The navigation accuracy of the inertial navigation system is mainly related to the zero-order drift coefficient and the first-order drift coefficient of the accelerometer and gyroscope. The values of the zero-order drift coefficient and the first-order drift coefficient are affected by many factors, such as the uncertainty of equipment manufacturing and changes in environmental conditions. Therefore, these drift coefficients can be regarded as random processes, and the values will change over time.

Therefore, the test data of this embodiment includes: four types of parameter sets, namely, the zero-order drift coefficient set of the accelerometer, the first-order drift coefficient set of the accelerometer, the zero-order drift coefficient set of the gyroscope and the first-order drift coefficient set of the gyroscope; each type of parameter set includes the drift coefficients corresponding to different directional axes; the drift coefficient corresponding to one directional axis is taken as a parameter.

Step 102: Using the Markov Monte Carlo method to interpolate missing values in the test data of the current period to obtain the periodic data of the current period, specifically including the following steps.

(1) The Metropolis-Hastings algorithm and the estimated potential scale reduction factor (EPSR) method are used to interpolate missing values for any feature in the test data of the current period to obtain the periodic data of each feature of the current period; a parameter in the test data is taken as a feature.

Specifically, for any feature in the test data of the current period, the process of interpolating missing values includes: constructing an initial Markov chain of the feature that obeys a stationary distribution according to the time series; determining the acceptance probability of state transition according to the proposed distribution and the stationary distribution; generating a random sample of the missing position according to the acceptance probability; inserting the random sample into the missing position of the initial Markov chain to generate a new Markov chain. The method of estimating the potential scale reduction factor is used to determine whether the new Markov chain converges. If converged, the random sample is inserted into the test data of the current period to obtain the periodic data of the feature in the current period; if not converged, the initial Markov chain of the feature that obeys a stationary distribution is reconstructed until a converged new Markov chain is generated.

(2) The periodic data of all characteristics of the current period are determined as the final periodic data of the current period.

Using the Metropolis-Hastings algorithm to obtain periodic data and using the complete data after periodic interpolation for model training can improve the model's predictive performance, but compared with using the original data, the interpolated samples may slightly reduce the interpretability and robustness of the model. In order to verify that the interpolated data can converge to the original data model, the method of estimating the potential scale reduction factor is also used to determine whether the sample converges to the probability distribution of the original data.

Step 103: Perform dimension reduction on the periodic data of the current period to obtain the periodic dimension reduction data of the current period.

Specifically, in high-dimensional data, there is often potential correlation between dimensions, resulting in large data redundancy and strong correlation. At the same time, the noise associated with high-dimensional data will reduce the efficiency and performance of data analysis and increase the complexity of problem analysis. One solution to this type of problem is to preprocess the data through PCA. PCA can effectively find the main direction of the data, remove dimensions that do not contain much information but carry noise, and greatly improve the prediction performance of the model. While retaining the main information of the data, PCA can eliminate the noise of high-dimensional data and reduce the dimension of the data as much as possible, which can avoid information loss while reducing the amount of calculation, greatly improving the efficiency of data analysis and the accuracy of the results.

Therefore, this embodiment uses PCA to reduce the dimension of the periodic data of the current period to obtain the periodic reduced dimension data of the current period.

Step 104: Input the periodic dimension reduction data of the current period into the health status prediction model to obtain the health status of the inertial navigation system in the current period.

Wherein, the health status prediction model is constructed based on machine learning methods.

In step 104, the method for determining the health status prediction model includes the following steps.

(1) Obtain the test data and corresponding health status of the inertial navigation system over the historical period.

(2) The missing values in the test data of the historical period are interpolated using the Markov Monte Carlo method to obtain the periodic data of the historical period. The interpolation process in this step is the same as the interpolation process in step 102 and will not be repeated here.

(3) Perform dimension reduction on the periodic data of the historical period to obtain the periodic dimension reduction data of the historical period. The dimension reduction method in this step is the same as the dimension reduction method in step 103, and will not be repeated here.

(4) Construct training data based on the periodic dimension reduction data of the historical period and the corresponding health status.

(5) Using the training data to train a support vector machine, and determining the trained support vector machine as the health status prediction model.

The support vector machine model is suitable for processing high-dimensional data with a small sample size. The multi-classification support vector machine is used to learn and predict the health status of the inertial navigation system. By classifying the health status of the inertial navigation system, the health status of the inertial navigation system at future moments is predicted. The established health status prediction model can be used to determine the health status of the inertial navigation system, and then evaluate the overall performance of the equipment, identify potential risks, and formulate corresponding preventive maintenance strategies.

The following focuses on further detailed introduction of the process of periodizing non-periodic data, using PCA to reduce data dimensionality, and establishing a health status prediction model.

1. Periodicize non-periodic data.

In the inertial navigation system, the zero-order drift coefficient and the first-order drift coefficient of the accelerometer and gyroscope in the non-periodic detection data x ^(t) (t=0,1,2...) can be approximated as a continuous random process. Since the value of x ^{(t +1)} in the random sequence is only related to x ^(t) , but not to other states in the random sequence. Therefore, the original non-periodic data can be periodized by the following steps.

1) Set each feature of the inertial navigation system to establish a Markov chain [x ⁽⁰⁾ , x ⁽¹⁾ , x ⁽²⁾ , ...] in time series. As time increases, a sufficiently long Markov chain is generated and follows a stable distribution. , x is a sample randomly sampled from the Markov chain.

2) Set the recommended distribution to , the proposed distribution means that the sampled values are from The probability of transitioning to x ^(t) , It is the state other than state x ^(t) in the Markov chain. The selection of the suggested distribution is determined subjectively by people according to the characteristics of the data distribution.

3) According to the stationary distribution , and get the acceptance probability The calculation formula for .

(1)

in, is the sampling value Obeying a stationary distribution, is the stationary distribution obeyed by the sampled value x ^(t) . is the sample value transferred from x ^(t) to The probability of Characterize the sample value from The probability of transitioning to x ^(t) .

4) Generate random numbers that follow an average distribution .

make, (2)

5) When the inertial navigation system has I features, assume that steps 2)-4) are looped C times to obtain C random samples, and each random sample is inserted into the initial Markov chain of each feature, and each feature generates C new Markov chains. Among them, The new Markov chain generated after inserting the sample into the original Markov chain of the i-th feature is the new Markov chain. There are S samples in represents the sth sample of the cth Markov chain in the ith feature, where . Use EPSR algorithm to determine the accepted samples After that, the model converges. The judgment process is as follows.

(1) Calculate the mean within the chain and the inter-chain mean .

(3)

(4)

(2) Calculate the variance B within each chain and the variance W between chains.

(5)

(6)

(3) Posterior variance Make an estimate.

(7)

(4) Calculation of EPSR value .

(8)

Generally, when EPSR < 1.1, the target is considered to have converged, and C random samples are interpolated into the original data set. Repeat steps 2) - 5) until the original data is periodized.

2. Use PCA to reduce data dimension.

1) Standardize the raw data.

Assume that the matrix X consists of N I-dimensional periodized sample data sets, and each row represents a set of sample data with I eigenvalues.

Right now, (9)

(10)

Normalize the matrix X to get the normalized matrix , The elements in are shown in formula (11) and formula (12).

(11)

(12)

In the formula, is the sample mean, S _i is the sample standard deviation, x _ni is the nth element in _Xi , for The nth element in .

2) Calculate eigenvalues and eigenvectors.

Let u be a unit vector. For ease of representation, the matrix The nth row of data is represented by x _n , and the matrix The nth row of data is inner-producted with u. When the N groups of sample data are inner-producted with u, the total variance is shown in formula (13).

(13)

in, is the covariance matrix of N groups of sample data, let is the covariance matrix If u is the eigenvalue of , then u is its corresponding eigenvector and T represents the transpose.

3) Select the principal components.

Assume that the I eigenvalues of the covariance matrix are as shown in formula (14).

(14)

Select the eigenvectors of the first M principal components whose cumulative variance contribution rate is greater than 85% according to the eigenvalue sorting. .

4) Calculate the principal components.

Eigenvector The principal components are calculated as a standard orthogonal basis, as shown in formula (15).

(15)

Where _Ym is the mth principal component (m=1,2,...,M), ( _um1 , _um2 ,..., _umi ) represents the standard orthogonal basis of the mth principal component, and the linear expression of the principal component is shown in formula (16).

(16)

3. Build a predictive model.

The reference levels of health status are "healthy", "sub-healthy" and "unhealthy". The multi-classification SVM algorithm is used to transform the multi-classification into several binary classifications. In this way, the data model with three classifications can be transformed into A binary classification model is created based on the reduced-dimensional data to perform classification prediction on new data.

After extracting the principal components, the data set has M samples, L of which are divided into training sets and the rest are used as test sets. The training samples are as shown in formula (17).

(17)

Where L is the number of samples, For the training samples, For the The classification labels of the training samples.

1) Construct the optimal hyperplane.

Assuming that the two categories of the binary classification model are p and q, the hyperplane equation is shown in formula (18).

(18)

After substituting all samples and their labels into formula (19), if the following conditions are met, the hyperplane can correctly classify the sample data. ω and b are parameters of the model, ω represents the weight, and b represents the bias.

(19)

Support Vectors is the heterogeneous sample that is closest to the hyperplane in the sample.

The distance d from the hyperplane is: (20)

2) Solve the classification function.

Establishing the objective function , and solve for the parameter ω.

(twenty one)

In formula (21), C is the penalty factor, is the slack variable, and Determines the allowable error of sample classification. For the The slack variables corresponding to the training samples are For the The Lagrange multiplier method is used to optimize equation (21) and obtain the new objective function Q(a).

(twenty two)

In formula (22), Q(a) represents the allowable error, For the The Lagrange multiplier corresponding to the training samples is: For the The Lagrange multiplier corresponding to the training samples is: For The number of training samples other than , K is the kernel function, For the training samples, For the The classification labels of training samples are obtained by solving the optimal values of parameters ω and b, determining the hyperplane equation, and thus establishing a prediction model.

An application example is given below to verify the effectiveness of the above-mentioned inertial navigation system health status prediction method.

The zero-order drift coefficient and the first-order drift coefficient of the gyroscope and accelerometer reflect the health status of the inertial navigation system. According to the actual situation, 37 sets of drift test data were collected, which had the problems of small sample size and unequal sampling intervals. The original data consists of the zero-order drift coefficient and the first-order drift coefficient of the accelerometer ( ), the zero-order drift coefficient and the first-order drift coefficient of the gyroscope ( ). Since the minimum sampling interval in the original data is one month, the data is interpolated as periodic detection data with an interval of one month. Among them, K _0X , K _0Y , K _0Z are the zero-order drift coefficients corresponding to the X-axis, Y-axis, and Z-axis of the accelerometer, K _1X , K _1Y , K _1Z are the first-order drift coefficients corresponding to the X-axis, Y-axis, and Z-axis of the accelerometer, D _0X , D _0Y , and D _0Z are the zero-order drift coefficients corresponding to the X-axis, Y-axis, and Z-axis of the gyroscope, and D _1X , D _1Y , and D _1Z are the first-order drift coefficients corresponding to the X-axis, Y-axis, and Z-axis of the gyroscope.

Step 1: Periodize the original non-periodic data.

Through the M-H algorithm and EPSR method, the original 37 sets of drift test data are interpolated into 50 sets of equal-periodic data.

Taking the test data of the zero-order drift coefficient of the gyroscope X-axis as an example, the interpolation effect is shown in Figure 2. As can be seen from Figure 2, the interpolated data (periodic data) is compared with the data before interpolation (original non-periodic data). The periodic data is consistent with the change trend of the original data. The statistical method t test is used to compare the similarity of the two groups of data under different indicators. The results are shown in Table 1.

Table 1 Difference analysis

Take the zero-order drift coefficient of the X-axis gyroscope as an example. A t-test was performed between the two sets of data before and after interpolation, and the results were h = 0 and p = 0.8002. According to the conventional practice in statistics, in general, when p > 0.05, the null hypothesis cannot be rejected, that is, there is no statistically significant difference between the two sets of data.

Step 2: Use PCA to reduce data dimension.

The principal components with cumulative variance contribution greater than 85% are selected as the first three principal components through the PCA algorithm. The 12-dimensional data is converted into 3-dimensional data. Since PCA dimensionality reduction is based on the theory of minimizing projection errors, it is a method to maximize variance. Whether the new data reduced by PCA contains the main information in the original data can support it to become the sample data for establishing a multi-classification support vector machine model.

Through calculation, the projection error after PCA dimensionality reduction is 4.768×10 ^-29 . The reconstruction error is close to zero, indicating that the data after dimensionality reduction is very close to the original data, and almost no information is lost. In addition, the dispersion of the values can be expressed by the variance of the data. As shown in Figure 3, the variance of the data after dimensionality reduction is analyzed as a proportion of the variance of the original data.

The variance proportion of each principal component after dimensionality reduction reflects the contribution of each principal component to the total variance. As shown in Figure 3, the variance proportion of the first three principal components is relatively large, and the data after dimensionality reduction can better retain the information of the original data. In summary, the variance proportion is large and the reconstruction error is small. It can be considered that the data after dimensionality reduction has the characteristics of preserving most of the information of the original data, being able to restore the structure and characteristics of the original data well, and using fewer dimensions to represent the key variance of the original data. These characteristics make the data after dimensionality reduction more compact and easy to process and analyze.

Step 3: Build a prediction model.

The PCA-reduced dataset is obtained through step 2, and 70% (35 groups) of the new data are randomly selected as training samples to train the prediction model based on the multi-classification support vector machine. The remaining 30% (15 groups) of data are used as test samples to evaluate the prediction effect of the prediction model. The original data without periodicity is predicted after dimensionality reduction and directly predicted to verify the effectiveness of the method.

In actual application, when performing a health status test on an inertial navigation system, it is more important to detect an ‘unhealthy’ state than a ‘healthy’ or ‘sub-healthy’ state. When testing the effectiveness of the prediction method, the positive sample is set to be a sample labeled ‘unhealthy’; the negative sample is a sample labeled ‘healthy’ or ‘sub-healthy’, indicating that the sample is not in a healthy state. Since the recall rate refers to the proportion of samples correctly predicted as positive samples by the model to all true positive samples, a high recall rate means that the model can better identify positive samples. F1-score combines accuracy and recall rate, taking into account both the model's ability to capture positive samples and the model's prediction accuracy for negative samples. Compared with accuracy and recall rate, F1-score can better comprehensively evaluate the performance of the classification model.

As shown in Table 2: When the inertial navigation system health status prediction method of the present invention is used, the accuracy of the model is 86.67%. This means that the proportion of samples predicted correctly by the model to the total number of samples is relatively high, and the prediction result is relatively accurate. The recall rate is 100.00%, indicating that the model can completely capture all real positive samples without missing any. The F1-score is 100.00%, indicating that the model performs well in the recognition of positive samples and the prediction accuracy of negative samples.

Table 2 Sample evaluation results

Embodiment 2

In order to execute the method corresponding to the above-mentioned embodiment 1 to achieve the corresponding functions and technical effects, an inertial navigation system health status prediction system is provided below.

Referring to Fig. 4, the system includes: a data acquisition module 201, which is used to acquire the test data of the current time period of the inertial navigation system; the test data is non-periodic data. An interpolation module 202, which is used to interpolate the missing values in the test data of the current time period using the Markov Monte Carlo method to obtain the periodic data of the current time period. A dimension reduction module 203, which is used to reduce the dimension of the periodic data of the current time period to obtain the periodic reduced dimension data of the current time period. A health status prediction module 204, which is used to input the periodic reduced dimension data of the current time period into the health status prediction model to obtain the health status of the inertial navigation system in the current time period. Wherein, the health status prediction model is constructed based on the machine learning method.

Embodiment 3

This embodiment provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the inertial navigation system health status prediction method of the first embodiment.

Optionally, the above-mentioned electronic device may be a server.

In addition, an embodiment of the present invention further provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the inertial navigation system health status prediction method of embodiment 1.

All of the above embodiments solve the limitations caused by limited test data and time interval uncertainty, and can improve the accuracy, stability and efficiency of inertial navigation system health status prediction, thereby positively affecting the reliability, performance and maintenance strategy of the equipment.

In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

The present specification uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims (6)

1. A method for predicting the health status of an inertial navigation system, comprising:

acquiring test data of the inertial navigation system in the current period; the test data are aperiodic data;

the method comprises the steps of interpolating missing values in test data of a current period by adopting a Markov Monte Carlo method to obtain periodic data of the current period;

performing dimension reduction on the periodic data of the current period to obtain periodic dimension reduction data of the current period;

inputting the periodic dimension reduction data of the current period into a health state prediction model to obtain the health state of the inertial navigation system in the current period;

wherein the health state prediction model is constructed based on a machine learning method;

the method for obtaining the periodic data of the current period by interpolating the missing value in the test data of the current period by adopting a Markov Monte Carlo method comprises the following steps:

adopting Mei Teluo wave Litsea-black Huntingth algorithm and a method for estimating potential scale reduction factors to interpolate missing values of any feature in test data of a current period to obtain periodic data of each feature of the current period; a parameter in the test data is used as a feature;

determining the periodic data of all the characteristics of the current period as the final periodic data of the current period;

the method for determining the health state prediction model comprises the following steps:

acquiring test data and corresponding health states of an inertial navigation system in a history period;

interpolating missing values in the test data of the historical period by adopting a Markov Monte Carlo method to obtain periodic data of the historical period;

performing dimensionality reduction on the periodic data of the historical period to obtain periodic dimensionality reduction data of the historical period;

constructing training data according to the periodic dimensionality reduction data of the historical period and the corresponding health state;

training a support vector machine by adopting the training data, and determining the trained support vector machine as the health state prediction model;

the method for obtaining the periodic data of each feature in the current period by interpolating the missing value of any feature in the test data in the current period by adopting a Mei Teluo Bolisi-Black-Ting algorithm and a method for estimating a potential scale reduction factor comprises the following steps:

for any feature in the test data of the current period, the process of interpolating the missing value includes:

constructing an initial Markov chain which obeys stable distribution of the features according to the time sequence;

determining the acceptance probability of the state transition according to the suggested distribution and the stable distribution;

generating a random sample of the missing position according to the acceptance probability;

inserting the random sample into the missing position of the initial Markov chain to generate a new Markov chain;

judging whether the Markov new chain is converged or not by adopting a method for estimating potential scale reduction factors;

if the random sample is converged, inserting the random sample into the test data of the current period to obtain the periodic data of the characteristics of the current period; if the characteristics are not converged, reconstructing an initial Markov chain which is compliant with stable distribution of the characteristics until a converged Markov new chain is generated.

2. The method for predicting the health status of an inertial navigation system according to claim 1, wherein the step of reducing the periodic data of the current period of time to obtain the periodic reduced data of the current period of time comprises:

and adopting a principal component analysis method to reduce the dimension of the periodic data in the current period to obtain the periodic dimension reduction data in the current period.

3. The inertial navigation system health prediction method of claim 1, wherein the test data comprises: the four parameter sets are a zero-order item drift coefficient set of the accelerometer, a primary item drift coefficient set of the accelerometer, a zero-order item drift coefficient set of the gyroscope and a primary item drift coefficient set of the gyroscope respectively; each type of parameter set comprises corresponding drift coefficients on different direction axes; the corresponding drift coefficient in one direction axis is used as a parameter.

4. An inertial navigation system health state prediction system, characterized in that it uses the inertial navigation system health state prediction method according to any one of claims 1 to 3, comprising:

the data acquisition module is used for acquiring test data of the inertial navigation system in the current period; the test data are aperiodic data;

the interpolation module is used for interpolating the missing value in the test data of the current period by adopting a Markov Monte Carlo method to obtain periodic data of the current period;

the dimension reduction module is used for reducing the dimension of the periodic data of the current period to obtain the periodic dimension reduction data of the current period;

the health state prediction module is used for inputting the periodic dimension reduction data of the current period into the health state prediction model to obtain the health state of the inertial navigation system in the current period;

wherein the health state prediction model is constructed based on a machine learning method.

5. An electronic device comprising a memory for storing a computer program and a processor that runs the computer program to cause the electronic device to perform the inertial navigation system health prediction method of any one of claims 1 to 3.

6. A computer readable storage medium, characterized in that it stores a computer program which, when executed by a processor, implements the inertial navigation system health state prediction method according to any one of claims 1 to 3.