KR102058686B1 - Design Method of the Terrain Referenced Navigation using Terrain Elevation Regression Model based Extreme Learning Machine and Computer readable storage medium having the same - Google Patents

Abstract

Provided is a topographical contour matching navigation design method capable of ensuring navigation in real-time while increasing learning accuracy of a rapid machine learning method. The topographical contour matching navigation design method comprises: (a) a step of deriving, by an earth elevation regression model module, a rapid machine learning-based earth elevation regression model in advance by using normalized learning data; and (b) a step of performing, by a topographical contour matching navigation performing system, topographical contour matching navigation with the rapid machine learning-based earth elevation regression model.

Description

Design method of the Terrain Referenced Navigation using Terrain Elevation Regression Model based Extreme Learning Machine and Computer readable storage medium having the same}

TECHNICAL FIELD The present invention relates to navigation technology, and more particularly to a generalized extreme learning machine based auto-encoder and / or a multi-layer radial basis function based artificial neural network. The present invention relates to a topographic navigation design method that can guarantee the real-time navigation of a navigation system while improving the learning accuracy of a rapid machine learning method using a network, and a computer-readable storage medium storing the method.

Currently, the Global Navigation Satellite System (GNSS) is used for precise navigation performance. However, there is a need for an alternative to obtain a stable navigation solution even in an environment where satellite navigation is impossible due to hostile radio disturbance.

In developed countries, research on Terrain Referenced Navigation (TRN) has been actively conducted as an alternative to guarantee the precision navigation performance in the GNSS unavailable environment. The Digital Elevation Model (DEM) is a map that shows the shape of the terrain by storing the altitude value of the terrain as a numerical value. A high resolution digital elevation model is necessary to improve the performance of the terrain control.

However, a large amount of memory space is required to store high-resolution digital elevation models. In addition, there is a delay in loading the numerical elevation model in a separate memory space into the navigation computer on which the terrain control navigation is performed, which hinders the real-time operation of the navigation.

In addition, various interpolation methods are used to obtain terrain altitude at a desired location using a non-continuous numerical elevation model, which also generates additional errors. As an alternative, the technology to derive the continuous Elevation Regression Model by learning the numerical elevation model by machine learning technique is an essential technique to solve the time delay problem caused by the dramatic saving of memory space and loading. .

However, there is a limitation that the real-time navigability of navigation should be at the same level or higher in terms of accuracy and accuracy of the existing method of calculating non-continuous high-resolution digital elevation model by interpolation as continuous surface elevation data.

  Considering the above conditions, the technique of designing a continuous surface elevation regression model through deep learning technique trained by Extreme Learning Machine (ELM) saves memory space and improves the performance of terrain control. And technology necessary for real-time guarantee of navigation.

Common techniques for designing machine-learning-based surface elevation regression models include B-spline based Artificial Neural Networks, Support Vector Machine Regression (SVMR), and single hidden layers. Classic rapid machine learning with a hidden layer-based artificial neural network.

However, these techniques have trained numerical elevation model data in a narrow area of hundreds of square meters and have not been able to derive higher accuracy than existing methods that match real numerical elevation models with interpolation. In addition, except for the B-spline-based artificial neural network, there is no case where the trained model is applied to the terrain control. In case of B-spline based neural network, the results of training low resolution digital elevation models in narrow areas are presented, which is not suitable for high performance terrain control.

In addition, in addition to the condition that learning must be performed with a very high accuracy in order to apply to the terrain control navigation, there is a restriction condition that the real time of navigation must be guaranteed. The regression model implemented by the B-spline-based artificial neural network and the supporter vector machine regression method requires a large amount of computation to reach a high level of accuracy, which makes it difficult to guarantee real-time. Compared to other technologies, the classical rapid machine learning technique is simple in structure and fast in learning, and can be obtained at a relatively high level of accuracy, and is currently used in various fields, but there are limitations in performance in complex nonlinear problems.

Despite the advantages of not needing a large amount of memory space as described above, machine learning regression models that do not have high learning accuracy that can replace numerical elevation models cannot be applied to terrain control. In addition, simulating a numerical elevation model with a complex model to increase accuracy can degrade the real-time navigation.

1. Korean Patent Publication No. 10-2010-0060194 2. Korean Patent Publication No. 10-2018-0052074

1. Tae Yoon Lee et al., "Automatic Generation of Digital Elevation Models with Improved Reliability Using High Resolution Satellite Images and Existing Digital Elevation Models", Korea Spatial Information Society, 2008

The present invention provides a generalized extreme learning machine based auto-encoder and / or a multi-layer radial basis function based artificial neural network in order to achieve the task presented above. It is an object of the present invention to provide a method for designing a terrain control navigation system capable of guaranteeing the real-time navigation of a navigation system while increasing the learning accuracy of a rapid machine learning method and a computer-readable storage medium storing the method.

The present invention provides a generalized extreme learning machine based auto-encoder and / or a multi-layer radial basis function based artificial neural network in order to achieve the task presented above. This paper provides a terrain control design method that can guarantee the real-time navigation of navigation while increasing the learning accuracy of rapid machine learning.

The terrain control design method,

(a) the surface elevation regression model module deriving a rapid machine learning based ground elevation regression model in advance using normalized training data; And

and (b) performing the terrain steering navigation system using the rapid machine learning-based ground elevation regression model.

In this case, the surface elevation regression model module includes: an input layer including first nodes formed of normalized latitude and longitude values of the normalized training data; First to third hidden layers, each consisting of second nodes using an activation function based on the first nodes; And an output layer for outputting the ground elevation regression model calculated using the second nodes.

In this case, the activation function is characterized in that the radiation basis function.

The first hidden layer is a rapid machine learning based oter encoder, and the rapid machine learning based otter encoder is an unsupervised learning method, and the output true value is the same as the input value.

및

(여기서,

은 m번째 학습 데이터의 위도값이며, λ^(m)는 m번째 학습 데이터의 경도 값이다. M은 전체 샘플 개수이며, φ와 Λ는 각각 M개의 위도, 경도 학습 데이터 입력이다)로 정의되는 것을 특징으로 한다.In addition, the first nodes are

And

(here,

Is the latitude value of the m th training data, and λ ^(m) is the longitude value of the m th training data. M is the total number of samples, and φ and Λ are M latitude and longitude learning data inputs, respectively).

In addition, the second nodes,

(여기서,

는 번째 노드의 활성화 함수이며, α_ki와 β _k는 k번째 노드에 할당된 방사 기저 함수의 평균값과 분산 값이고,

은 유클리드 거리(Euclidean distance)를 의미하며, K는 은닉층의 총 노드 수를 나타낸다)으로 이루어지는 것을 특징으로 한다,Equation

(here,

Is the activation function of the th node, α _ki and β _k are the mean and variance of the radiated basis function assigned to the k th node,

Is the Euclidean distance, and K represents the total number of nodes of the hidden layer.

In addition, the mean value and the variance of the basis of the radiation basis is characterized in that it is determined through k-mean clustering (K-mean clustering).

The second hidden layer and the third hidden layer are rapid machine learning based artificial neural networks, and the rapid machine learning based cited neural network is a supervised learning method. It is characterized by the).

In addition, the terrain control navigation system is characterized by updating the measurement of the terrain information using the rapid machine learning-based ground elevation regression model and the difference value for the difference between the measured value between the altimeter and barometric altitude.

On the other hand, another embodiment of the present invention stores a computer readable storage medium storing program code for executing the method of terrain control navigation using the rapid machine learning based ground elevation regression model described above.

According to the present invention, a rapid machine learning method using a generalized extreme learning machine based auto-encoder and / or a multi-layer radial basis function based artificial neural network It can guarantee the real-time navigability of the navigation while increasing the learning accuracy of.

In addition, another effect of the present invention is that since it simulates more precisely in a rugged area, it works more advantageously on the measurement effective for the terrain control navigation.

1 is a conceptual diagram showing the structure of the surface elevation regression module for generating a rapid machine learning-based ground elevation regression model according to an embodiment of the present invention.
FIG. 2 is a flowchart illustrating a learning procedure according to the rapid machine learning based regression model shown in FIG. 1.
3 is a block diagram of a system for performing terrain control navigation using a rapid machine learning based ground elevation regression model according to an embodiment of the present invention.
4 is a graph of experimental results according to the linear interpolation method which is widely used in general terrain control.
5 is a graph of experimental results according to the rapid machine learning-based ground elevation regression model according to an embodiment of the present invention.
6 is an experimental result comparing a general linear interpolation method and a regression model based on rapid machine learning according to an embodiment of the present invention.
7 to 10 are experimental results comparing the integrated navigation performance of the BKF-based terrain control navigation and inertial navigation system navigation information according to FIG.

As the invention allows for various changes and numerous embodiments, particular embodiments will be illustrated in the drawings and described in detail in the written description. However, this is not intended to limit the present invention to specific embodiments, it should be understood to include all modifications, equivalents, and substitutes included in the spirit and scope of the present invention.

In describing each drawing, like reference numerals are used for like elements. Terms such as first and second may be used to describe various components, but the components should not be limited by the terms. The terms are used only for the purpose of distinguishing one component from another.

For example, without departing from the scope of the present invention, the first component may be referred to as the second component, and similarly, the second component may also be referred to as the first component. The term “and / or” includes any combination of a plurality of related items or any item of a plurality of related items.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art.

Terms such as those defined in the commonly used dictionaries should be construed as having meanings consistent with the meanings in the context of the related art, and are not construed in ideal or excessively formal meanings unless expressly defined in this application. Should not.

Hereinafter, a method for designing a terrain control navigation using a surface elevation regression model based on rapid machine learning and a computer readable storage medium storing the method will be described in detail with reference to the accompanying drawings.

In general, a digital elevation model is a map representing three-dimensional data of a terrain by storing the altitude value of the terrain at a given resolution interval. Bilinear interpolation is generally used to obtain terrain altitude at a desired location using a discontinuous numerical elevation model. This is expressed as the following equation.

Where f (x, y) is the terrain altitude at the desired location (x, y), and h (Q ₁₁ ), h (Q ₁₂ ), h (Q ₂₁ ) and h (Q ₂₂ ) are each at the desired location Q is the closest numerical elevation data. Q is the longitude and latitude value of the desired position as input data of the numerical elevation data.

The linear interpolation method calculates the desired terrain altitude by reflecting the weight inversely proportional to the linear distance between the numerical elevation data stored at the grid interval and the desired position.

To perform terrain control based on this linear interpolation method, a large memory is needed to store the digital elevation model.

An embodiment of the present invention proposes a technique for deriving a rapid machine learning-based surface elevation regression model using numerical elevation model data in advance and performing a terrain control method using such a rapid machine learning-based surface elevation regression model.

Particularly, in one embodiment of the present invention, a generalized ELM based auto-encoder and a multi-layered radiated basis function are man-made in consideration of significant memory space savings, improved terrain control performance, and guarantee of real-time navigation. We design a regression model using a neural network (Multi-layer Radial Basis Function based Artificial Neural Network) and propose a method to perform terrain control.

1 is a conceptual diagram showing the structure of the surface elevation regression module 100 for generating a rapid machine learning-based ground elevation regression model according to an embodiment of the present invention. Prior to explaining FIG. 1, when describing rapid machine learning, rapid machine learning is described in G.-B. First introduced by Huang, deep learning is widely used in various fields because of its superior performance and fast learning speed compared to backpropagation learning.

Classic rapid machine learning is a front artificial artificial neural network with a single hidden layer. The input weight and unit bias of the hidden layer are initially set randomly once and are no longer adjusted. On the other hand, the output weight of the hidden layer is analytically calculated by the least square method.

Artificial neural networks designed by the classical rapid machine learning are not applicable to the fields that need to process large data or big data and that require high accuracy, compared to the advantages of easy to implement and fast learning speed.

To solve this problem, FIG. 1 shows a modified rapid machine learning concept consisting of three hidden layers. Referring to FIG. 1, the ground elevation regression model module 100 includes an input layer 110, a first hidden layer 120-1, a second hidden layer 120-2, a third hidden layer 120-3, and an output layer. 150 and the like.

The input layer 110 is composed of nodes (x ₁ , x ₂ ) composed of normalized learning data latitude / longitude, and is represented by the following equation.

은 m번째 학습 데이터의 위도값이며, λ^(m)는 m번째 학습 데이터의 경도 값이다. M은 전체 샘플 개수이며, φ와 Λ는 각각 M개의 위도, 경도 학습 데이터 입력이다. 학습 데이터는 보유하고 있는 수치표고자료 데이터로부터 획득할 수 있다. 즉, 학습 데이터의 입력은 정규화된 위, 경도 값이며, 학습 참값은 정규화된 수치표고도 값이다.here,

Is the latitude value of the m th training data, and λ ^(m) is the longitude value of the m th training data. M is the total number of samples, and φ and Λ are M latitude and longitude learning data inputs, respectively. The training data can be obtained from the numerical elevation data held. That is, the input of the training data is the normalized latitude and longitude values, and the learning true value is the normalized numerical elevation value.

The hidden layers 120-1 to 120-3 are composed of nodes that use a radiated basis function as an activation function.

는 번째 은닉 노드의 활성화 함수이며, α_ki와 β_k는 k번째 은닉 노드에 할당된 방사 기저 함수의 평균과 분산 값이다. 또한,

은 유클리드 거리(Euclidean distance)를 의미하며, K는 은닉층의 총 노드 수를 나타낸다. 각 은닉층 방사 기저 함수의 평균값 및 분산값은 k 평균 군집(K-mean clustering) 등의 비지도 학습을 통해 결정된다.here,

Is the activation function of the th hidden node, and α _ki and β _k are the mean and variance of the radiated basis function assigned to the k th hidden node. Also,

Is the Euclidean distance, and K is the total number of nodes in the hidden layer. The mean and variance of each hidden-layer radiative basis function are determined through unsupervised learning, such as k-mean clustering.

The first hidden layer 120-1 is a rapid machine learning based otter encoder, and the second and third hidden layers 120-2 and 120-3 are formed of a general rapid machine learning based artificial neural network. Rapid machine learning based OTC encoder is unsupervised learning method, and output true value is same as input value, while general rapid machine learning based artificial neural network is supervised learning method, and output true value is a preset target value included in training data. (target value). Rapid machine learning based artificial neural network may be a multi-layer Radial Basis Function based Artificial Neural Network.

The output H ₁ of the hidden layer nodes is expressed as a matrix as follows.

Assuming that the output weight matrix of the first hidden layer 121 is C ₁ , the output weight matrix C ₁ can be calculated by the least square method in consideration of the characteristic that the output of the Otter encoder is the same as the input. have.

는 무어-펜로즈 일반화된 역행렬(Moore-Penrose generalized inverse)이며, 이는 일반적으로 특이값 분해(Singular value composition)로 계산된다. λ는 과적합(Overfitting) 방지를 위한 리지 규정 파라미터(Ridge regulation parameter)이며, I는 항등 행렬(Identity matrix)이고, T는 전치행렬을 나타낸다. 또한,

이며,

이다. here,

Is a Moore-Penrose generalized inverse, which is generally calculated by singular value composition. λ is a ridge regulation parameter for preventing overfitting, I is an identity matrix, and T is a transpose matrix. Also,

Is,

to be.

Through the output weight matrix C ₁ of the first hidden layer 120-1 learned by the above equation, the OTC encoder mediation vector 130 is calculated as in the following equation.

이며, 이는 제 2 은닉층(120-2)의 입력이 된다. 제 2 은닉층(120-2)의 방사 기저 함수는 위에서 기술한 바와 같이 활성화 함수로 방사 기저 함수를 사용하는 노드(node)들로 구성되고 이는 수식 4와 같이 동일한 방식으로 결정된다. At this time,

This is an input of the second hidden layer 120-2. The radiation basis function of the second hidden layer 120-2 is composed of nodes that use the radiation basis function as an activation function as described above, and is determined in the same manner as in Equation 4.

Assuming that the output weight matrix of the second hidden layer 120-2 is C ₂ , the generalized rapid mechanics method is calculated by the following equation.

와 같고, 행렬요소인 d^(m)는 수치표고모델의 지형 고도를 정규화한 데이터이다. At this time, the target value matrix of the training data is

The matrix element d ^(m) is the normalized terrain elevation of the digital elevation model.

The mediation vector 140 learned through the above equation is calculated as follows.

이며, 이는 제 3 은닉층(120-3)의 입력이 된다. 제 3 은닉층(120-3)의 방사 기저 함수는 위에서 기술한 바와 같이 활성화 함수로 방사 기저 함수를 사용하는 노드(node)들로 구성되고 이는 수식 4와 같이 동일한 방식으로 결정된다.At this time,

This is the input of the third hidden layer 120-3. The radiation basis function of the third hidden layer 120-3 is composed of nodes that use the radiation basis function as an activation function as described above, and is determined in the same manner as in Equation 4.

는 제안한 방사기저 함수 기반 제 3은닉층 모델을 통해 산출되는 행렬이다. 획득된 학습 데이터로부터

는 수학식 8의 행렬 연산에서

대신

로 변경하여 계산할 수 있다. 따라서 수학식 9에서

와

대신 산출된

와

의 행렬곱으로 출력층(y)이 계산된다(y).The output layer 150 outputs the surface elevation regression model y according to the above.

Is a matrix calculated from the proposed radiative basis function based third hidden layer model. From the acquired training data

In the matrix operation

instead

Can be changed to So in equation (9)

Wow

Calculated instead

Wow

The output layer y is calculated by the matrix product of (y).

The auto-encoder mediation vector 130 and mediation vector 140 shown in FIG. 1 are vectors for explaining a learning method, and are not actual hidden layers. A total of three hidden layers are provided, and the first hidden layer 120-1 serves to distribute more radiation basis functions in samples of more severe terrain of the numerical elevation model samples. The second hidden layer 120-2 and the third hidden layer 120-3 play a role of improving the learning accuracy of the ground elevation regression model.

FIG. 2 is a flowchart illustrating a learning procedure according to the rapid machine learning based regression model shown in FIG. 1. Referring to FIG. 2, training data normalization is performed (step S210). In other words, the input layer 110 is composed of nodes composed of normalized latitude and longitude values of the training data. Afterwards, the hidden layers 120-1 to 120-3 are composed of nodes that use a radiation basis function as an activation function, and k-mean clustering in the first hidden layer 120-1. The average value and the variance value of the radiated basis function are determined through unsupervised learning such as (step S220).

Thereafter, the output weight matrix is calculated by the least square method in consideration of the characteristic that the output of the Otter encoder is the same as the input (step S230). An output of the first hidden layer 120-1 becomes an input of the second hidden layer 120-2.

Thereafter, in the second hidden layer 120-2, the average value and the variance of the radiated basis function are determined through the rapid machine learning based quoting neural network, and the output weighting matrix is calculated by the least square method (steps S240 and S250). An output of the second hidden layer 120-2 becomes an input of the third hidden layer 120-3.

Thereafter, in the third hidden layer 120-3, the average value and the variance of the radiated basis function are determined through the rapid machine learning based artificial neural network, and the output weighting matrix is calculated by the least square method (steps S260 and S270).

In this way, by training off-line using the training data and implementing the surface elevation regression model through software, it is possible to perform terrain control navigation without installing a separate numerical elevation model in a large memory.

3 is a block diagram illustrating a system for performing a terrain control navigation system using a rapid machine learning-based ground elevation regression model according to an embodiment of the present invention. Referring to FIG. 3, when learning is performed according to FIGS. 1 and 2, the ground elevation regression model y is generated by the ground elevation regression model module 100.

The terrain control navigation system 300 includes an inertial navigation device 301 for generating navigation information, an altimeter 302 for generating relative distance measurement information to an indicator closest to an antibody, and a barometric altimeter for generating barometric pressure measurement information. 303, the terrain control navigation module 310 for receiving navigation information, the navigation information, altitude measurement information, and the integrated navigation execution module 320 for receiving the barometric altitude measurement information may be configured.

The altimeter 302 may use a radio altimeter or an interferometer radar altimeter. The integrated navigation execution module 320 executes integrated navigation combining the terrain control navigation and inertial navigation information. The terrain control navigation filter 322 for performing the terrain control navigation performs time propagation by the second time propagation performing module 321 using the inertial navigation device navigation information. The measurement information of the altimeter 302 is determined by the measurement validity determination module 325 as valid for the terrain control navigation, and when determined to be valid, the terrain control filter 322 updates the measurement value by the measurement update module 326. If it is determined that it is not valid, only time propagation is performed.

The difference in measurement between the altimeter 302 and the barometric altimeter 303 calculated by the first subtractor 324-1 is the terrain altitude measured at the current position of the antibody, and is calculated by the second subtractor 324-2. The difference from the rapid machine learning based ground elevation regression model y is used as a measure of the terrain control filter 322. In addition, the terrain validity determination module 323 determines whether the terrain roughness or uniqueness is valid through the residual verification of the terrain steering filter 322 in the measurement update module 326, and then passes the result to the terrain steering navigation execution module 310. It is used for the measurement update of the measurement update module 313 configured. When the terrain control execution module 310 does not satisfy the terrain validity determination condition, the first time propagation performing module 311 performs only time propagation by the navigation navigation filter 312. When the terrain validity determination condition is satisfied, the integrated navigation filter measurement update module 313 performs the update by using the latitude, longitude, and altitude estimates, which are outputs of the measurement update module 326, as measurement values.

The term “… module” described in the drawings refers to a unit that processes at least one function or operation, which may be implemented in software and / or hardware. In hardware implementations, application specific integrated circuits (ASICs), digital signal processing (DSP), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, microprocessors, and the like are designed to perform the functions described above. It may be implemented in an electronic unit or a combination thereof. In the software implementation, the module may be implemented as a module that performs the above-described function. The software may be stored in a memory unit and executed by a processor. The memory unit or processor may employ various means well known to those skilled in the art.

Figure 4 is a graph of the experimental results according to the linear interpolation method that is widely used in the general terrain control method, Figure 5 is a graph of the experimental results according to the surface elevation regression model based on rapid machine learning according to an embodiment of the present invention.

4 and 5, the rapid machine learning-based ground elevation regression proposed by the present invention using 5,391,738 (161 MB) of sample data of an area at DTED (Digital Terrain Elevation Data) level 3 (resolution 10m) of the numerical elevation model. This is a result of comparing the simulation accuracy between the model 510 and the linear interpolation method 410 which is currently used in the terrain control.

From this, we can see that they are almost equivalent, which means that the rapid machine learning based ground elevation regression model can be applied to the terrain control. In addition, when using the rapid machine learning-based surface elevation regression model, memory space savings of 1/1533 times can be expected to be 105 KB to store learning model parameters instead of 161 MB.

This difference is larger when the sample data is larger, and the regression model parameters learned in the reserved memory space of the navigation computer for terrain control can be stored. On the other hand, when linear interpolation is used, a large amount of space may be required, and thus a time delay for loading may occur. Therefore, the proposed invention may also expect a time delay error and / or a development cost saving effect.

6 is an experimental result comparing a general linear interpolation method and a regression model based on rapid machine learning according to an embodiment of the present invention. Referring to FIG. 6, the proposed invention compares the integrated navigation performance of the BKF (Bank-of-Kalman Filter) based terrain control and inertial navigation system navigation information with the existing linear interpolation based performance (610). .

7 to 10 are experimental results (710, 810, 910, 1010) comparing the integrated navigation performance of the BKF-based terrain control navigation and inertial navigation system navigation information according to FIG. Referring to FIGS. 7 to 10, it can be seen that the terrain control method improves the proposed method by about 2.514 mCEP and the integrated navigation method improves about 1.217 mCEP. The linear interpolation and the learning accuracy of the present invention are equal, but the reason why the terrain control and integrated navigation using the same improves the performance is due to the large numerical elevation model processing error in the rough area with large nonlinearity. . On the contrary, the rough area is more precisely simulated by the first hidden layer (120-1 in FIG. 1) that includes the otter encoder proposed in the present invention, so that it is more advantageous for the effective measurement for the terrain control navigation. . Therefore, the proposed invention can be expected to improve the performance of terrain control.

In addition, the steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied in the form of program instructions that may be executed by various computer means such as a microprocessor, a processor, a central processing unit (CPU), or the like, for reading a computer It can be recorded on a possible medium. The computer readable medium may include program (instruction) code, data file, data structure, etc. alone or in combination.

The program (command) code recorded on the medium may be those specially designed and configured for the present invention, or may be known and available to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs, DVDs, Blu-rays, etc., and ROMs and RAMs. Semiconductor memory devices specifically configured to store and execute program (command) code, such as RAM), flash memory, and the like.

Here, examples of program (instruction) code include not only machine code generated by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.

100: surface elevation regression model module
110: input layer
120-1 to 120-3: first to third hidden layers
150: output layer
300: terrain control navigation system

Claims (10)

(a) the surface elevation regression model module 100 deriving a rapid machine learning based ground elevation regression model y in advance using normalized training data; And
(b) performing a terrain steering navigation system using the rapid machine learning based ground elevation regression model (y);
Terrain control navigation design method using a rapid machine learning-based surface elevation regression model comprising a.
The method of claim 1,
The surface elevation regression model module 100,
An input layer (110) consisting of first nodes (x ₁ , x ₂ ) composed of normalized latitude and longitude values among the normalized learning data;
First to third hidden layers 120-1, 120-2, and 120-3 consisting of second nodes 121 using an activation function based on the first nodes (x ₁ , x ₂ ); And
Terrain control navigation design method using a rapid machine learning-based ground elevation regression model comprising a; output layer (150) for outputting the ground elevation regression model calculated using the second nodes (121).
The method of claim 2,
And wherein the activation function is a radial basis function.
The method of claim 3, wherein
The first hidden layer 120-1 is a rapid machine learning based otter encoder, and the rapid machine learning based otter encoder is an unsupervised learning method, and the output true value is the same as the input value. Terrain control design method using regression model.
The method of claim 3, wherein
The first nodes (x1, x2) is the equation

And

(here,

Is the latitude value of the m th training data, and λ ^(m) is the longitude value of the m th training data. M is the total number of samples, φ and Λ are the input of M latitude and longitude learning data, respectively).
The method of claim 5,
The second nodes,
Equation

(here,

Is the activation function of the th node, α _ki and β _k are the mean and variance of the radiative basis function assigned to the k th node,

Is a Euclidean distance, and K represents the total number of nodes in the hidden layer.
The method of claim 6,
The mean and variance values of the basis of the radiation basis are determined using k-mean clustering. The method of claim 2,
The second hidden layer 120-2 and the third hidden layer 120-3 are rapid machine learning based artificial neural networks, and the rapid machine learning based artificial neural networks are supervised learning methods, and output true values are included in normalized learning data. A method for terrain control navigation using a rapid machine learning-based surface elevation regression model, which is a predetermined target value.
The method of claim 1,
The terrain control navigation system 300 updates the measurement information of the terrain information by using the rapid machine learning-based ground elevation regression model y and the difference value of the difference between the measured values between the altimeter 302 and the barometric altimeter 303. Terrain control method design using rapid machine learning based ground elevation regression model.
A computer-readable storage medium storing program code for executing the method of terrain control navigation using the rapid machine learning based ground elevation regression model according to any one of claims 1 to 9.

Images