Elman neural network assisting tight-integrated navigation method without GNSS signals

Patent metadata

Field	Value
Publication number	US-11821729-B2
Application number	US-202017013750-A
Country	US
Kind code	B2
Filing date	Sep 7, 2020
Priority date	Sep 26, 2019
Publication date	Nov 21, 2023
Grant date	Nov 21, 2023

Abstract

A tight-integrated navigation method assisted by Elman neural network when GNSS signals are blocked based on the tight-integrated navigation system model of the INS and GNSS, where the dynamic Elman neural network prediction model is used to train the inertial navigation error model and the GNSS compensation model, so as to solve the problem of tight-integrated navigation when the GNSS signals are blocked. When the GNSS signals are blocked, the trained neural network is used to predict the output error of GNSS and compensate the output of inertial navigation, so that the error will not diverge sharply, and the system can continue to work in the integrated navigation mode. The low-cost tight-integrated navigation module is used, and the collected information is preprocessed to form the sample data for training the neural network to train the Elman neural network model.

First claim

What is claimed is:

1. An Elman neural network assisting tight-integrated navigation method without Global navigation satellite system (GNSS) signals of an integrated navigation system, comprising:
step 1: establishing an Elman neural network and selecting a hidden layer transfer function of the Elman neural network,
step 2: designing an Elman learning algorithm,
step 3: establishing a tight-integrated Kalman filter and detecting an availability of the GNSS signals,
step 4: in response to detecting the availability of the GNSS signals, navigating a ship using the GNSS signals and causing a neural network to work in a training mode, wherein a three-dimensional position information output by a receiver of an inertial navigation system is used as input samples for training the neural network designed in the step 2 by a processor; a compensation value of an inertial navigation error output after a fusion of the Kalman filter established in step 3 is used as expected output samples for training the network, and the input samples and the expected output samples are brought into the Elman neural network established in the step 2 for training the Elman neural network;

when errors between actual outputs of the Elman neural network and the expected output samples are more than a predetermined threshold, an updated value of a network weight is obtained by cyclically using the Elman neural network algorithm designed in step 2 until the errors between the actual outputs of the network and the expected outputs are less than the predetermined threshold, and
step 5: in response to detecting that the GNSS signals are blocked, causing the neural network to work in a prediction mode; a navigation position information output by the inertial navigation system is used as the input of the network trained by the step 4, an error of the inertial navigation system is predicted by the neural network trained by the step 4, and a corrected navigation information is obtained by correcting the navigation output of the inertial navigation system by the error predicted by the neural network; wherein the navigation output of the inertial navigation system is corrected and compensated by the predicted data, thus the integrated navigation system is configured to continue to provide navigation data when the GNSS signals are momentarily unavailable; and
navigating the ship using the corrected and compensated output to the integrated navigation system of a ship to meet navigation accuracy requirements of the ship;
wherein, the Elman neural network comprises an input layer, a hidden layer, a connection layer and an output layer, a mathematical model of the Elman neural network is expressed as follows:
x h ( k )= f ( W 1 P ( k )+ W 3 x ( k ));
x c ( k )=α x h ( k− 1);

y ( k )= g ( W 2 x ( k ));

wherein, P(k) represents an input vector of the Elman neural network at time k, xn(k) represents an output vector of the hidden layer neuron at time k, x(k) represents an input vector of the Elman neural network derived from the output of the hidden layer at time k, x c (k) represents an output vector of the connection layer at time k, y(k) represents the output vector of the entire network output layer at time k, W 1 , W 2 and W 3 are respectively connection weight matrixes between the input layer and the hidden layer, between the hidden layer and the output layer, and between the connection layer and the hidden layer, f(⋅) and g(⋅) are respectively transfer functions of the hidden layer and the output layer, and a is a connection feedback gain factor; and
wherein the inertial navigation system comprises a gyroscope and an accelerometer.

2. The method as claimed in claim 1 , wherein, an S-tangent function is selected as the transfer function of the hidden layer of Elman neural network.

3. The method as claimed in claim 1 , wherein, the step 2 comprises,
a calculation process of the Elman neural network is divided to comprise a forward propagation of a working signal and a backward propagation of the error;
the calculation of the forward propagation of the working signal is in consistent with the mathematical model of the Elman neural network, and the signal y(k) of the network output is calculated;
the back propagation of the error is as follows: that the actual output of the network is y(k) at time k is assumed, and the expected output response of the network is y d (k), then the error of the network is as follows:
E k =½( y d ( k )− y ( k )) 2

partial derivatives of the error function with respect to the connection weights between different layers are obtained, respectively; the partial derivative of the error function E k with respect to the connection weight matrix W 2 from the hidden layer to the output layer is obtained as follows:

∂
E

∂

w
ij
2

=

-

(

y

d
,
i

(
k
)

-

y
⁡
(
k
)

)

⁢

∂

y
i

(
k
)

∂

w
ij
2

=

-

(

y

d
,
i

(
k
)

-

y
⁡
(
k
)

)

⁢

g
⁡
(
•
)

⁢

x
j

(
k
)

Assignees

• Univ Harbin Eng

Inventors

• Zhao Lin
• Peng Zihang
• Ding Jicheng
• Wang Kun
• Bai Yaguo
• Zhang Yongchao
• Wang Renlong

Classifications

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G01C21/188 Primary

  for accumulated errors, e.g. by coupling inertial systems with absolute positioning systems · CPC title

• G01C21/165 Primary

  combined with non-inertial navigation instruments · CPC title

• G01S19/47

  the supplementary measurement being an inertial measurement, e.g. tightly coupled inertial · CPC title