阅读说明：本技术 一种基于高斯过程回归的信息更新频率提升方法 (Information updating frequency improving method based on Gaussian process regression ) 是由 孙骞 王娱萤 许前昆 叶方 李一兵 田园 于 2021-08-03 设计创作，主要内容包括：本发明提出一种基于高斯过程回归的信息更新频率提升方法,所述方法首先利用通用卷积谱混合成分核函数(GeneralizedConvolutionS pectralMixtureKernel,GCSMK)方法建立原子干涉陀螺仪(AtomInterferenceGyroscope,AIG)惯性测量数据的结构化模型,获得惯性测量信息之间的相互依赖关系；然后建立基于GCSMK的高斯过程回归(GaussianProcessesRegression,GPR)算法框架,利用基于稀疏化思想改进的GPR算法实现惯性传感器信息更新频率的快速提升,从而在不改变惯性传感器自身精度的前提下提升惯性信息的更新频率。(The invention provides an information updating frequency improving method based on Gaussian process regression, which comprises the steps of firstly, establishing a structural model of Atomic Interference Gyroscope (AIG) inertia measurement data by using a general convolution spectrum mixed component kernel (GCSMK) method, and obtaining the mutual dependency relationship between inertia measurement information; and then establishing a Gaussian Process Regression (GPR) algorithm framework based on GCSMK, and realizing quick increase of the information updating frequency of the inertial sensor by utilizing a GPR algorithm improved based on a sparsification thought, so that the updating frequency of the inertial information is increased on the premise of not changing the self precision of the inertial sensor.)

1. An information updating frequency improving method based on Gaussian process regression is characterized in that: the method comprises the following steps:

step 1: an atomic interference gyroscope AIG inertial sensor is arranged on equipment or a carrier, and is preheated and used for acquiring output data of an AIG;

step 2: determining a kernel function form of a general convolution spectrum mixed component kernel function GCSMK, setting an initial value of a hyper-parameter, and determining a prior model of a regression GPR in a Gaussian process in a probability distribution form according to output data of an AIG;

and step 3: training by using historical information of the AIG inertial sensor as a sample, and solving an optimal hyper-parameter according to a sparsification thought;

and 4, step 4: obtaining a posterior model, namely a sparse GCSMK-GPR prediction model IGCSMK-GPR, according to the prior model and the optimal hyper-parameter;

and 5: and predicting the output of the AIG inertial sensor by using an IGCSLK-GPR prediction model, so as to obtain the frequency-boosted information of the AIG inertial sensor.

2. The method of claim 1, wherein: in step 3, a part of samples in the AIG historical samples are selected and initialized to a pseudo sample setThen simultaneously completing the AIG pseudo sample set by a conjugate gradient methodAnd optimization of hyper-parameters.

3. The method of claim 2, wherein: in the step 5, the process is carried out,

1) if the AIG inertial sensor at the current sampling moment has no measurement information, performing high-precision prediction on the output of the AIG inertial sensor by using an IGCSCK-GPR prediction model, and taking prediction data as the output of the current sampling moment;

2) if the AIG inertial sensor at the current sampling moment has the measurement information, the measurement information is output at the current moment, and the prediction model is corrected by using the measurement information, so that the accuracy of the prediction information of the AIG inertial sensor is improved.

4. The method of claim 3, wherein: the kernel function form of the generic convolution spectrum mixed component kernel function GCSMK specifically comprises the following steps:

wherein Q is the number of kernel function components, τ is time, μ_ijIs a cross mean, sigma_ijFor cross covariance, θ_ijFor cross-over time delay, phi_ijIs the cross phase; c. C_ijThe dependence between different components in GCSMK can be measured as a constant term independent of τ.

5. The method of claim 4, wherein: in step 2, assuming that the time delay and the phase are both 0, the weight w is set_iMean value of μ_iSum variance Σ_iDetermining a prior model of the GCSMK-GPR in a probability distribution form of a Bayesian Gaussian mixture model according to initial values of the hyperparameters:

where f is a prior model function.

6. The method of claim 5, wherein: the sparse GCSMK-GPR prediction model comprises the following steps:

wherein f is^*In order to predict the function of the model,andare all optimal hyperparameters.

Technical Field

The invention relates to the technical field of inertial navigation, in particular to a method for improving the information updating frequency of an inertial sensor based on Gaussian Process Regression (GPR), which is more specifically to utilize a general Convolution spectrum mixed component Kernel function (GCSMK) method to obtain the interdependence relation between inertial sensor data and utilize a GPR algorithm improved based on a sparseness idea to realize the intelligent and rapid improvement of the information updating frequency of the inertial sensor under the framework of the method, thereby improving the updating frequency of the inertial sensor on the premise of not changing the self precision of the inertial sensor.

Background

The performance of the gyroscope, which is used as a core sensor of the inertial navigation system, is the most important factor influencing the accuracy and robustness of the inertial navigation system. In recent decades, under the traction of urgent needs of national defense construction, the high-precision gyroscope technology has been developed rapidly, and the technology has been developed from a first generation rotor gyroscope based on Newton mechanics to a second generation optical gyroscope based on wave optics. The static gyroscope is used as a rotor gyroscope with the highest precision and is widely arranged in a strategic nuclear submarine, but the dynamic measurement precision of the static gyroscope is seriously influenced by the problems of rotor drift, rotation speed attenuation, long stabilization time and the like caused by the limitation of a beryllium rotor processing technology, and the further development of an inertial navigation system of the static gyroscope is restricted; as a representative of the optical gyroscope, the optical fiber gyroscope and the laser gyroscope are also increasingly applied to tactical weapons and submarines, however, in a multi-external disturbance coupling environment, the internal stress field of the optical fiber gyroscope changes, which results in unstable dynamic scale factors, the self-machine shake of the laser gyroscope is coupled with the external disturbance, which makes the structural mode of the shock absorber change and the shock absorption mode difficult to optimize, high-precision measurement in the multi-external disturbance coupling environment cannot be ensured, and the requirement of a future inertial navigation system is difficult to meet. With the implementation of important strategies such as the continuous advancement of China to the deep and distant sea and the strong ocean, the requirements on the performance of a navigation system are higher and higher, and the research on the inertial navigation technology based on a novel ultrahigh-precision gyroscope is urgently needed.

With the emergence of the third nobel prize in the quantum field, the rapid development of the technologies such as Atom excitation, Atom capture, quantum state superposition, and atomic group temperature reduction, the atomic Gyroscope inertial navigation system based on quantum mechanics becomes a research hotspot in the foreign navigation field, wherein an Atomic Interference Gyroscope (AIG) based on the atomic de broglie wave Interference principle is a Gyroscope with the highest precision potential at present, and is becoming the focus of the research on the ultrahigh-precision inertial navigation system.

In recent years, the technology of the latent AIG inertial navigation system makes certain breakthrough. However, no matter the space domain AIG or the time domain AIG, in order to ensure the inertia measurement accuracy of the AIG, the atomic capture and cooling process needs to be extended, so that the update frequency of the measurement information of the AIG is low, all dynamic information of the submarine cannot be accurately captured, and particularly, in a complex environment, the AIG inertia measurement with low information update frequency causes a large navigation calculation error, and the requirement of high-precision dynamic navigation cannot be met. Therefore, how to increase the update frequency of the inertial information without changing the accuracy of the AIG device is one of the hot spots in current research.

The currently common data frequency boosting methods are generally methods such as a spatial data interpolation method and a multi-sensor combination method. The spatial data interpolation method generally uses a mathematical model to fit historical data and predict future short-term data so as to realize the frequency improvement, and mainly comprises methods such as polynomial interpolation, linear interpolation, quadratic interpolation, Newton interpolation, spline interpolation and the like; however, the method generally needs a model with data known in advance, which is difficult to satisfy in a practical environment, so that the method has great limitation in use. The multi-sensor combination method is that a system is combined by using sensors with multiple sampling frequencies, so that the data updating frequency of the whole system is improved; the cost required for this approach is significantly increased and the performance of the introduced sensor also affects the measurement performance of the combined system.

Disclosure of Invention

The invention provides an information updating frequency improving method based on Gaussian process regression for improving the frequency and the precision of output signals of AIG inertial sensors by considering the nonlinear non-stationarity characteristic of AIG measurement information in an actual system. The invention utilizes a general Convolution spectrum mixed component Kernel (GCSMK) method to obtain the interdependence relation between the inertial sensor data, and utilizes a GPR algorithm improved based on a sparsification idea to realize the intelligent and rapid promotion of the information updating frequency of the inertial sensor under the framework of the method, thereby promoting the updating frequency of the inertial information on the premise of not changing the self precision of the inertial sensor.

The invention is realized by the following technical scheme, and provides an information updating frequency improving method based on Gaussian process regression, which comprises the following steps:

step 1: an atomic interference gyroscope AIG inertial sensor is arranged on equipment or a carrier, and is preheated and used for acquiring output data of an AIG;

step 2: determining a kernel function form of a general convolution spectrum mixed component kernel function GCSMK, setting an initial value of a hyper-parameter, and determining a prior model of a regression GPR in a Gaussian process in a probability distribution form according to output data of an AIG;

and step 3: training by using historical information of the AIG inertial sensor as a sample, and solving an optimal hyper-parameter according to a sparsification thought;

and 4, step 4: obtaining a posterior model, namely a sparse GCSMK-GPR prediction model IGCSMK-GPR, according to the prior model and the optimal hyper-parameter;

and 5: and predicting the output of the AIG inertial sensor by using an IGCSLK-GPR prediction model, so as to obtain the frequency-boosted information of the AIG inertial sensor.

Further, in step 3, a portion of the samples in the AIG history sample is selected and initialized to a pseudo sample setThen simultaneously completing the AIG pseudo sample set by a conjugate gradient methodAnd optimization of hyper-parameters.

Further, in step 5,

1) if the AIG inertial sensor at the current sampling moment has no measurement information, performing high-precision prediction on the output of the AIG inertial sensor by using an IGCSCK-GPR prediction model, and taking prediction data as the output of the current sampling moment;

2) if the AIG inertial sensor at the current sampling moment has the measurement information, the measurement information is output at the current moment, and the prediction model is corrected by using the measurement information, so that the accuracy of the prediction information of the AIG inertial sensor is improved.

Further, the kernel function form of the generic convolution spectrum mixed component kernel function GCSMK specifically includes:

wherein Q is the number of kernel function components, τ is time, μ_ijIs a cross mean, sigma_ijFor cross covariance, θ_ijFor cross-over time delay, phi_ijIs the cross phase; c. C_ijThe dependence between different components in GCSMK can be measured as a constant term independent of τ.

Further, in step 2, assuming that the time delay and the phase are both 0, the weight w is set_iMean value of μ_iSum variance Σ_iDetermining a prior model of the GCSMK-GPR in a probability distribution form of a Bayesian Gaussian mixture model according to initial values of the hyperparameters:

where f is a prior model function.

Further, the sparse GCSMK-GPR prediction model is as follows:

wherein f is^*In order to predict the function of the model,andare all optimal hyperparameters.

The invention has the advantages that: according to the invention, the output signal of the inertial sensor is processed by using an information updating frequency improving method based on IGCSCK-GPR, so that the characteristics that the adaptive machine learning technology can carry out adaptive modeling on data and predict unseen trend are utilized, and the better super-parameter interpretability and the optional advantage of model optimization are realized by combining a Gaussian process regression algorithm, and the precision and the robustness of inertial sensor prediction and frequency increase are improved; in addition, the invention also introduces a sparsification thought, further reduces the time complexity of the algorithm and improves the frequency increasing efficiency of the sensor.

Drawings

FIG. 1 is a schematic flow diagram of the process of the present invention;

FIG. 2 is a graph of training set data output signal results;

FIG. 3 is a graph of output signal results after processing using the present invention;

fig. 4 is a graph of the error between the output signal and the true value after processing using the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The invention provides a GPR (Improved GCSMK GPR, IGCSLK-GPR) information updating frequency improving method based on Improved GCSMK. Firstly, establishing a structural model of AIG inertial measurement data by using a GCSMK method to obtain the interdependence relation between inertial measurement information; and then establishing a GPR algorithm framework based on GCSMK, and realizing quick improvement of the information updating frequency of the inertial sensor by utilizing the GPR algorithm improved based on the sparsification thought, so that the updating frequency of the inertial information is improved on the premise of not changing the self precision of the inertial sensor.

With reference to fig. 1-4, the present invention provides an information updating frequency improving method based on gaussian process regression, which includes the following steps:

step 1: an Atomic Interference Gyroscope (AIG) inertial sensor is arranged on equipment or a carrier, and is preheated and used for collecting output data of the AIG;

step 2: determining a Kernel function form of a General Convolution Spectral Mixture Kernel (GCSMK), setting an initial value of a hyper-parameter, and determining a prior model of Gaussian Process Regression (GPR) in a probability distribution form according to output data of the AIG;

the kernel function form of the generic convolution spectrum mixed component kernel function GCSMK specifically comprises the following steps:

wherein Q is the number of kernel function components, τ is time, μ_ijIs a cross mean, sigma_ijFor cross covariance, θ_ijFor cross-over time delay, phi_ijIs the cross phase; c. C_ijThe dependence between different components in GCSMK can be measured as a constant term independent of τ.

In step 2, without loss of generality, assuming that both the delay and the phase are 0, the weight w is set_iMean value of μ_iSum variance Σ_iDetermining a prior model of the GCSMK-GPR in a probability distribution form of a Bayesian Gaussian mixture model according to initial values of the hyperparameters:

where f is a prior model function.

And step 3: training by using historical information of the AIG inertial sensor as a sample, and solving an optimal hyper-parameter according to a sparsification thought;

in step 3, a part of samples in the AIG historical samples are selected and initialized to a pseudo sample setThen simultaneously completing the AIG pseudo sample set by a conjugate gradient methodAnd optimizing the hyper-parameters to obtain the optimal hyper-parametersAnd

and 4, step 4: obtaining a posterior model, namely a sparse GCSMK-GPR prediction model (Improved GCSMK GPR, IGCSLK-GPR), according to the prior model and the optimal hyper-parameter;

the sparse GCSMK-GPR prediction model comprises the following steps:

wherein f is^*In order to predict the function of the model,andare all optimal hyperparameters.

And 5: and predicting the output of the AIG inertial sensor by using an IGCSLK-GPR prediction model, so as to obtain the frequency-boosted information of the AIG inertial sensor.

In the step 5, the process is carried out,

1) if the AIG inertial sensor at the current sampling moment has no measurement information, performing high-precision prediction on the output of the AIG inertial sensor by using an IGCSCK-GPR prediction model, and taking prediction data as the output of the current sampling moment;

2) if the AIG inertial sensor at the current sampling moment has the measurement information, the measurement information is output at the current moment, and the prediction model is corrected by using the measurement information, so that the accuracy of the prediction information of the AIG inertial sensor is improved.

In this way, the frequency increase of updating the AIG information based on the IGCSLK-GPR prediction model can be completed.

The effect of the invention can be verified by the following tests:

firstly, a simulation test environment is set up, the sampling frequency of the AIG is set to be 1Hz, and the effect of the invention is verified by using AIG output data. The number of kernel function components in a sample is 3, the initial value of a mean value in a hyper-parameter is 0, the initial value of a variance in the hyper-parameter is [1, 0.6, 1] respectively, 500 data in historical data are selected as a sample set, then the optimization of the hyper-parameter is completed by using a conjugate gradient method to obtain a posterior GCSMK-GPR prediction model, and finally the short-term prediction is carried out on the data of the AIG to obtain the output information in the future short term, so that the frequency promotion of the AIG information is realized.

The simulated AIG original signal is shown in fig. 2. The signal obtained after 5s of prediction is obtained by performing frequency up processing on the original data by using the information updating frequency increasing method provided by the invention is shown in fig. 3, and the difference between the predicted value and the true value is shown in fig. 4.

As can be seen from fig. 3 and 4, the data output in the future 5s can be well predicted by using the present invention, the prediction accuracy gradually decreases with the time delay, and the prediction errors at the 1 st s and the 5 th s are 5.63% and 15.58%, respectively. When performing frequency lifting, only the data of the future 1s need to be predicted. Therefore, the invention provides an effective solution for solving the frequency increase problem of the inertial sensor, and effectively improves the dynamic performance and the applicability of the inertial navigation system.

The information updating frequency improving method based on gaussian process regression proposed by the present invention is described in detail above, and the principle and the implementation manner of the present invention are explained in this document by applying specific examples, and the description of the above examples is only used to help understanding the method and the core idea of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, there may be variations in the specific embodiments and the application scope, and in summary, the content of the present specification should not be construed as a limitation to the present invention.