CN103196446A - Intelligent gyro signal filtering method of trapdown inertial navigation system in high overload environment - Google Patents

Abstract

The invention discloses an intelligent gyro signal filtering method of a trapdown inertial navigation system in a high overload environment. The intelligent gyro signal filtering method comprises the following steps: establishing three BP networks to simulate gyro outputs on axes X, Y and Z in the high overload environment, and directly installing a gyro and an accelerometer of the trapdown inertial navigation system on a carrier; judging whether the carrier is in the high overload environment or not according to the output of the accelerometer; if the output of the accelerometer is larger than a set threshold, judging that the carrier enters the high overload environment; if not, judging that the carrier is not in the high overload environment; carrying out the navigation calculation with the outputs of the three gyros when the carrier is not in the high overload environment, and taking the outputs of the three gyros as online training samples of the BP networks to ensure that network parameters are consistent with a current carrier movement state; and enabling the three BP networks operating in the analog output state to simulate gyro signal output when the carrier is judged to be in the high overload environment, and thereby ensuring the stable operation of the trapdown inertial navigation system.

Description

Strapdown inertial navigation system gyroscope signal intelligent filtering method under the high overload environment

Technical field

The present invention relates generally to field of navigation technology, especially designs the strapdown inertial navigation system gyroscope signal intelligent filtering method under a kind of high overload environment.

Background technology

Strapdown inertial navigation system has the advantage that volume is little, in light weight, cost is low, anti-overload ability is strong, is widely used in fields such as aviation and navigation at present, and in the high overload application, strapdown inertial navigation system has special advantages especially.The measuring accuracy of the key sensor as strapdown inertial navigation system--gyro has directly determined performance index and the precision index of total system, is the goal in research of inertial technology so improve the measuring accuracy of gyro always.Be subjected to the restriction of gyro structural design, production level, measurement bandwidth and other working mechanism, under the high overload environment, accelerometer output can reach 5g～50g usually, distortion can appear in gyro output signal inevitably, therefore reduce that the distortion of gyro output signal is an important gordian technique in inertial technology field to the influence of strapdown inertial navigation system precision always under the high overload environment.

Summary of the invention

Goal of the invention: at the problem of gyro output signal appearance distortion under the high overload environment in the object peculiar to vessel, the present invention proposes a kind of based on the strapdown inertial navigation system gyroscope signal intelligent filtering method under the high overload environment of BP network technology.

The present invention utilizes the BP network method for real-time learning, modeling and the simulation characteristics of non-linear variable, realizes gyro signal artificial intelligence output function under the high overload environment by increase a strapdown inertial navigation system gyroscope signal intelligent filtering dedicated thread with the parallel running of strapdown inertial navigation system navigation calculating main thread in the computing machine of strapdown inertial navigation system.The present invention has under non-overload environment, navigation is calculated main thread and is finished navigation calculating by gathering the actual output of gyro, gyro signal intelligent filter thread then is operated in learning state, utilize current gyro output to upgrade BP network internal structural parameters, better guarantee BP network internal parameter and current carrier movement state consistency; When being under the high overload environment, control gyro signal intelligent filter thread switches to the simulation output state, navigation is calculated main thread and is adopted gyro simulation output to replace the actual output of gyro, has avoided the measuring accuracy influence of gyro output signal distortion to strapdown inertial navigation system like this; After high overload finished, gyro signal artificial intelligence module returned to online training study pattern again.The present invention is applicable to the strapdown inertial navigation system peculiar to vessel under the high overload environment, has satisfied navigational system under the high overload environment, and namely accelerometer is output as under 5g～50g situation, the demand of held stationary output.

Technical scheme of the present invention is specific as follows:

Strapdown inertial navigation system gyroscope signal intelligent filtering method under a kind of high overload environment:

Step 1: set up 3 BP network structures, described 3 BP network initial configurations comprise the BP network of Z axle gyro output under the BP network of X-axis gyro output under the simulation high overload environment, the BP network of simulating Y-axis gyro output under the high overload environment and the simulation high overload environment, 3 BP network initial configurations are: the three-layer network version of single input/single output and single hidden layer, the hidden layer node number is m, m=10; The input of 3 BP networks is respectively the gyro output time correlation x on X, Y, the Z axle ^θ, x ^γ, x

, output is respectively the gyro output y on X, Y, the Z axle ^θ, y ^γ, y

, wherein x, y represent the input of BP network, output respectively, subscript θ, γ,

Represent inputing or outputing of corresponding X, Y, Z axle respectively, the BP network structure as shown in Figure 3; The output of simulation X-axis gyro output BP network output layer is found the solution synoptic diagram as shown in Figure 4, and simulation Y, the output of Z axle gyro output BP network output layer are found the solution synoptic diagram and the output of simulation X-axis gyro output BP network output layer to find the solution synoptic diagram similar; The output of simulation X-axis gyro output BP network hidden layer is found the solution synoptic diagram as shown in Figure 5, synoptic diagram and the output of simulation X-axis gyro output BP network output layer are found the solution in the output of simulation Y, Z axle gyro output BP network hidden layer, and to find the solution synoptic diagram similar, and the output of 3 BP networks is respectively with relation between importing:

$y^{\mathrm{\θ}}=g^{\mathrm{\θ}}\left[\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\θ}}{f}_j^{\mathrm{\θ}}\left(w_{\mathrm{ij}}^{\mathrm{\θ}}{x}^{\mathrm{\θ}}\right)\right]$

$y^{\mathrm{\γ}}=g^{\mathrm{\γ}}\left[\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\γ}}{f}_j^{\mathrm{\γ}}\left(w_{\mathrm{ij}}^{\mathrm{\γ}}{x}^{\mathrm{\γ}}\right)\right]$

In the formula

Represent 3 BP network input layers respectively to the connection weights of hidden layer, subscript i, j represent i node of input layer and j node of hidden layer respectively; Represent 3 BP network hidden layers respectively to the connection weights between the output layer, subscript k represents k node of output layer, i=1, k=1;

Representing 3 BP network hidden layer activation functions respectively, all is the Sigmoidal function, and the Sigmoidal function expression is as follows:

$f_j^{\mathrm{\θ}}\left(u_j^{\mathrm{\θ}}\right)=\frac{1}{1+e^{{-u}_j^{\mathrm{\θ}}}}$

$f_j^{\mathrm{\γ}}\left(u_j^{\mathrm{\γ}}\right)=\frac{1}{1+e^{{-u}_j^{\mathrm{\γ}}}}$

Its derivative is:

$f_j^{\mathrm{\θ}\′}\left(u_j^{\mathrm{\θ}}\right)=\frac{1}{1+e^{{-u}_j^{\mathrm{\θ}}}}\frac{e^{{-u}_j^{\mathrm{\θ}}}}{1+e^{{-u}_j^{\mathrm{\θ}}}}=[1-f_j^{\mathrm{\θ}}\left(u_j^{\mathrm{\θ}}\right)]f_j^{\mathrm{\θ}}\left(u_j^{\mathrm{\θ}}\right)$

$f_j^{\mathrm{\γ}\′}\left(u_j^{\mathrm{\γ}}\right)=\frac{1}{1+e^{{-u}_j^{\mathrm{\γ}}}}\frac{e^{{-u}_j^{\mathrm{\γ}}}}{1+e^{{-u}_j^{\mathrm{\γ}}}}=[1-f_j^{\mathrm{\γ}}\left(u_j^{\mathrm{\γ}}\right)]f_j^{\mathrm{\γ}}\left(u_j^{\mathrm{\γ}}\right)$

Wherein,

Represent the input of 3 BP network hidden layer activation functions respectively, and

Be exponential function;

Represent the output of 3 BP network hidden layer activation functions respectively,

g ^θ(), g ^γ(), Representing 3 BP network output layer activation functions respectively, all is the Sigmoidal function, and the Sigmoidal function expression is as follows:

$g^{\mathrm{\θ}}\left(u^{\′\mathrm{\θ}}\right)=\frac{1}{1+e^{{-u}^{\′\mathrm{\θ}}}}$

$g^{\mathrm{\γ}}\left(u^{\′\mathrm{\γ}}\right)=\frac{1}{1+e^{{-u}^{\′\mathrm{\γ}}}}$

Its derivative is:

$g^{\mathrm{\θ}\′}\left(u^{\′\mathrm{\θ}}\right)=\frac{1}{1+e^{{-u}^{\′\mathrm{\θ}}}}\frac{e^{{-u}^{\′\mathrm{\θ}}}}{1+e^{{-u}^{\′\mathrm{\θ}}}}=[1-g^{\mathrm{\θ}}\left(u^{\′\mathrm{\θ}}\right)]g^{\mathrm{\θ}}\left(u^{\′\mathrm{\θ}}\right)$

$g^{\mathrm{\γ}\′}\left(u^{\′\mathrm{\γ}}\right)=\frac{1}{1+e^{{-u}^{\′\mathrm{\γ}}}}\frac{e^{{-u}^{\′\mathrm{\γ}}}}{1+e^{{-u}^{\′\mathrm{\γ}}}}=[1-g^{\mathrm{\γ}}\left(u^{\′\mathrm{\γ}}\right)]g^{\mathrm{\γ}}\left(u^{\′\mathrm{\γ}}\right)$

Wherein, u ' ^θ, u ' ^γ,

Represent the input of 3 BP network output layer activation functions respectively, and $u^{\′\mathrm{\θ}}=\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\θ}}{f}_j^{\mathrm{\θ}}\left(w_{\mathrm{ij}}^{\mathrm{\θ}}{x}^{\mathrm{\θ}}\right),$ $u^{\′\mathrm{\γ}}=\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\γ}}{f}_j^{\mathrm{\γ}}\left(w_{\mathrm{ij}}^{\mathrm{\γ}}{x}^{\mathrm{\γ}}\right),$

Step 2: receive the gyro signal on X, Y, the Z axle respectively;

Step 3: judge whether to be in and enter the high overload environment, described high overload environment determination methods is: the accelerometer output according to strapdown inertial navigation system is judged, when accelerometer output is acceleration of gravity greater than setting threshold 5g and g, then be considered as entering the high overload environment, otherwise be considered as non-high overload environment;

Step 4: when judged result when entering the high overload environment, then 3 BP networks are operated in the simulation output state, simulation gyro signal output, and respectively with the output y of current up-to-date step 1 ^θ, y ^γ,

Calculate as the simulating signal output of gyro on X, Y, the Z axle and for navigation;

Step 5: when judged result is non-high overload environment, then 3 gyro outputs are used for navigation and calculate and the gyro under the non-high overload environment is exported as the online training sample of BP network, described online training sample comprises gyro output time correlation With the gyro output under the non-high overload environment

Gyro accumulative total hits is p, makes p=p+1 again, if p is less than online number of training P, P=300 then returns step 2, if p equals 300, then utilize 300 training samples that the BP network is carried out online training, and with current online training result step of updating 1 described BP network connection weights

Return step 2 after the p zero clearing, described online training comprises the online training of X-axis, the online training of Y-axis and the online training of Z axle;

The step of the online training of described X-axis is as follows:

Step 5.1.1: make iterations n _θInitial value be 1, utilize random function Random () that the BP network hidden layer of X-axis gyro output under the simulation high overload environment and the weights of output layer are carried out initialization:

$w_{\mathrm{ij}}^{\mathrm{\θ}}=\begin{array}{cc}\mathrm{Random}(\·)& w_{\mathrm{ij}}^{\mathrm{\θ}}\∈\left[\mathrm{0,1}\right]\end{array}$

$w_{\mathrm{jk}}^{\mathrm{\θ}}=\begin{array}{cc}\mathrm{Random}(\·)& w_{\mathrm{jk}}^{\mathrm{\θ}}\∈\left[\mathrm{0,1}\right]\end{array}$

Step 5.1.2: calculate used 300 sample total errors in the current iteration process

$E_{A\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}{E}_{p\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=\frac{1}{2}\underset{p=1}{\overset{300}{\mathrm{\Σ}}}{(d_p^{\mathrm{\θ}}-y_{p\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}})}^2$

In the formula

Be the output of the BP network of X-axis gyro output under the simulation high overload environment when importing p sample in the current iteration process;

$y_{p\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=g^{\mathrm{\θ}}\left[\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\θ}}{f}_j^{\mathrm{\θ}}\left(w_{\mathrm{ij}}^{\mathrm{\θ}}{x}_p^{\mathrm{\θ}}\right)\right]$

Step 5.1.3: if Perhaps n _θ〉=10000, then adopt current

Under the simulation high overload environment of step of updating 1 in the BP network of X-axis gyro output

Otherwise enter step 5.1.4;

Step 5.1.4: make n _θ=n _θ+ 1, $w_{\mathrm{ij}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}+\mathrm{\Δ}{w}_{\mathrm{ij}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}},w_{\mathrm{jk}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}+{\mathrm{\Δw}}_{\mathrm{jk}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}$

Wherein

$\mathrm{\Δ}{w}_{\mathrm{ij}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=-0.58\frac{\∂E_A^{\mathrm{\θ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}$

$\mathrm{\Δ}{w}_{\mathrm{jk}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=-0.58\frac{\∂E_A^{\mathrm{\θ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}$

Wherein

$\frac{\∂E_A^{\mathrm{\θ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\θ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\underset{j=1}{\overset{10}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\θ}}}{\∂y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}\frac{\∂y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}{\∂u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}\frac{\∂u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}{\∂x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}\frac{\∂x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}{\∂u_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}\frac{\∂u_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}$

$=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\underset{j=1}{\overset{10}{\mathrm{\Σ}}}(d_p^{\mathrm{\θ}}-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}(1-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}{x}_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}(1-x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}})x_p^{\mathrm{\θ}}$

$\frac{\∂E_A^{\mathrm{\θ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\θ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\θ}}}{\∂y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}\frac{\∂y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}{\∂u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}\frac{\∂u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}(d_p^{\mathrm{\θ}}-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})g^{\mathrm{\θ}\′}\left(u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}\right)x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}$

$=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}(d_p^{\mathrm{\θ}}-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}(1-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}$

In the formula

Represent the BP network output when p sample of input in the last iterative process, Represent the input of the BP network output layer activation function when p sample of input in the last iterative process,

Represent the output of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent the input of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent that last iteration finishes the back hidden layer to the connection weights of output layer,

For last iteration finishes the back input layer to the connection weights of hidden layer, return step 5.1.2;

The step of the online training of described Y-axis is as follows:

Step 5.2.1: make iterations n _γInitial value be 1, utilize random function Random () that the BP network hidden layer of Y-axis gyro output under the simulation high overload environment and the weights of output layer are carried out initialization:

$w_{\mathrm{ij}}^{\mathrm{\γ}}=\begin{array}{cc}\mathrm{Random}(\·)& w_{\mathrm{ij}}^{\mathrm{\γ}}\∈\left[\mathrm{0,1}\right]\end{array}$

$w_{\mathrm{jk}}^{\mathrm{\γ}}=\begin{array}{cc}\mathrm{Random}(\·)& w_{\mathrm{jk}}^{\mathrm{\γ}}\∈\left[\mathrm{0,1}\right]\end{array}$

Step 5.2.2: calculate used 300 sample total errors in the current iteration process

$E_{A\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}{E}_{p\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=\frac{1}{2}\underset{p=1}{\overset{300}{\mathrm{\Σ}}}{(d_p^{\mathrm{\γ}}-y_{p\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}})}^2$

In the formula

Be the output of the BP network of Y-axis gyro output under the simulation high overload environment when importing p sample in the current iteration process;

$y_{p\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=g^{\mathrm{\γ}}\left[\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\γ}}{f}_j^{\mathrm{\γ}}\left(w_{\mathrm{ij}}^{\mathrm{\γ}}{x}_p^{\mathrm{\γ}}\right)\right]$

Step 5.2.3: if

Perhaps n _γ〉=10000, then adopt current

Under the simulation high overload environment of step of updating 1 in the BP network of Y-axis gyro output

Otherwise enter step 5.2.4;

Step 5.2.4: make n _γ=n _γ+ 1, $w_{\mathrm{ij}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}+\mathrm{\Δ}{w}_{\mathrm{ij}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}},$ $w_{\mathrm{jk}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}+\mathrm{\Δ}{w}_{\mathrm{jk}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}$

Wherein

$\mathrm{\Δ}{w}_{\mathrm{ij}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=-0.58\frac{\∂E_A^{\mathrm{\γ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}$

$\mathrm{\Δ}{w}_{\mathrm{jk}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=-0.58\frac{\∂E_A^{\mathrm{\γ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}$

Wherein

$\frac{\∂E_A^{\mathrm{\γ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\γ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\underset{j=1}{\overset{10}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\γ}}}{\∂y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}\frac{\∂y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}{\∂u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}\frac{\∂u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}{\∂x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}\frac{\∂x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}{\∂u_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}\frac{\∂u_{\mathrm{jp}}^{\mathrm{\γ}}(n_{\mathrm{\γ}}-1)}{\∂w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}$

$=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\underset{j=1}{\overset{10}{\mathrm{\Σ}}}(d_p^{\mathrm{\γ}}-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}(1-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}{x}_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}(1-x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}})x_p^{\mathrm{\γ}}$

$\frac{\∂E_A^{\mathrm{\γ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\γ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\γ}}}{\∂y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}\frac{\∂y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}{\∂u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}\frac{\∂u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}(d_p^{\mathrm{\γ}}-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})g^{\mathrm{\γ}\′}\left(u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}\right)x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}$

$=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}(d_p^{\mathrm{\γ}}-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}(1-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}$

In the formula

Represent the BP network output when p sample of input in the last iterative process,

Represent the input of the BP network output layer activation function when p sample of input in the last iterative process,

Represent the output of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent the input of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent that last iteration finishes the back hidden layer to the connection weights of output layer,

For last iteration finishes the back input layer to the connection weights of hidden layer, return step 5.2.2;

The step of the online training of described Z axle is as follows:

Step 5.3.1: make iterations Initial value be 1, utilize random function Random () that the BP network hidden layer of Z axle gyro output under the simulation high overload environment and the weights of output layer are carried out initialization:

Step 5.3.2: calculate used 300 sample total errors in the current iteration process

In the formula

Be the output of the BP network of Z axle gyro output under the simulation high overload environment when importing p sample in the current iteration process;

Step 5.3.3: if

Perhaps

Then adopt current Under the simulation high overload environment of step of updating 1 in the BP network of Z axle gyro output

Otherwise enter step 5.3.4;

Step 5.3.4: order

Wherein

Wherein

In the formula

Represent the BP network output when p sample of input in the last iterative process,

Represent the input of the BP network output layer activation function when p sample of input in the last iterative process,

Represent the output of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent the input of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent that last iteration finishes the back hidden layer to the connection weights of output layer,

For last iteration finishes the back input layer to the connection weights of hidden layer, return step 5.3.2.

The relative carrier mounting structure with accelerometer of the gyro of strapdown inertial navigation system such as Fig. 1, strapdown inertial navigation system gyroscope signal intelligent filtering fundamental diagram such as Fig. 2.

Compared with prior art, the present invention has following advantage:

1) solved the distortion of under high overload environment gyro output signal to the influence of system accuracy, guaranteed the strapdown inertial navigation system smooth working, and the present invention be software approach, do not need system hardware is made any modification that the historical facts or anecdotes border implements to make things convenient for, feasible;

2) based on the long-term work experience and repeatedly the experiment method, optimize network structure, determine the training sample number, guaranteed that the gyro signal intelligent filter method algorithm based on the BP network technology that adopts is simple, real-time, not strong for the Computing Capability Requirement, be convenient to Project Realization;

3) working method that adopts the sub-thread parallel operation of navigation calculating main thread and gyro signal intelligent filter, the sub-thread of gyro signal intelligent filter to learn afterwards to simulate earlier, under non-overload environment, the BP network is in the training study state all the time, it is consistent with the motion state of current boats and ships that its inner parameter keeps, therefore when switching to simulation output services state, simulation gyro output signal that can be more true to nature satisfies and switches flexible, the reliable requirement of navigational system stable working.

Description of drawings

Fig. 1 is the relative carrier mounting structure with accelerometer of the gyro figure of strapdown inertial navigation system of the present invention.

Fig. 2 is strapdown inertial navigation system gyroscope signal intelligent filtering fundamental diagram of the present invention.

Fig. 3 is BP network structure of the present invention.

Fig. 4 finds the solution synoptic diagram for the present invention simulates the output of X-axis gyro output BP network output layer.

Fig. 5 finds the solution synoptic diagram for the present invention simulates the output of X-axis gyro output BP network hidden layer.

Fig. 6 is embodiment of the invention simulation X-axis gyro output BP network training result.

Fig. 7 is embodiment of the invention simulation X-axis gyro output BP network test result.

Fig. 8 is embodiment of the invention gyro signal artificial intelligence module application theory diagram.

Fig. 9 resolves the result for the actual output of embodiment of the invention X-axis gyro.

Figure 10 resolves the result for the auxiliary X-axis output of embodiment of the invention gyro signal artificial intelligence module.

Figure 11 resolves figure as a result for the auxiliary output contrast of the actual output of embodiment of the invention X-axis gyro and gyro signal artificial intelligence module.

Embodiment

Below in conjunction with specific embodiment, further illustrate the present invention, should understand these embodiment only is used for explanation the present invention and is not used in and limits the scope of the invention, after having read the present invention, those of ordinary skills all fall within the application's claims institute restricted portion to the modification of the various equivalent form of values of the present invention.

Strapdown inertial navigation system gyroscope signal intelligent filtering method under a kind of high overload environment:

Step 1: set up 3 BP network structures, described 3 BP network initial configurations comprise the BP network of Z axle gyro output under the BP network of X-axis gyro output under the simulation high overload environment, the BP network of simulating Y-axis gyro output under the high overload environment and the simulation high overload environment, 3 BP network initial configurations are: the three-layer network version of single input/single output and single hidden layer, the hidden layer node number is m, m=10; The input of 3 BP networks is respectively the gyro output time correlation x on X, Y, the Z axle ^θ, x ^γ,

Output is respectively the gyro output y on X, Y, the Z axle ^θ, y ^γ,

Wherein x, y represent the input of BP network, output respectively, subscript θ, γ,

Represent inputing or outputing of corresponding X, Y, Z axle respectively, the BP network structure as shown in Figure 3; The output of simulation X-axis gyro output BP network output layer is found the solution synoptic diagram as shown in Figure 4, and simulation Y, the output of Z axle gyro output BP network output layer are found the solution synoptic diagram and the output of simulation X-axis gyro output BP network output layer to find the solution synoptic diagram similar; The output of simulation X-axis gyro output BP network hidden layer is found the solution synoptic diagram as shown in Figure 5, synoptic diagram and the output of simulation X-axis gyro output BP network output layer are found the solution in the output of simulation Y, Z axle gyro output BP network hidden layer, and to find the solution synoptic diagram similar, and the output of 3 BP networks is respectively with relation between importing:

$y^{\mathrm{\θ}}=g^{\mathrm{\θ}}\left[\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\θ}}{f}_j^{\mathrm{\θ}}\left(w_{\mathrm{ij}}^{\mathrm{\θ}}{x}^{\mathrm{\θ}}\right)\right]$

$y^{\mathrm{\γ}}=g^{\mathrm{\γ}}\left[\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\γ}}{f}_j^{\mathrm{\γ}}\left(w_{\mathrm{ij}}^{\mathrm{\γ}}{x}^{\mathrm{\γ}}\right)\right]$

In the formula Represent 3 BP network input layers respectively to the connection weights of hidden layer, subscript i, j represent i node of input layer and j node of hidden layer respectively;

Represent 3 BP network hidden layers respectively to the connection weights between the output layer, subscript k represents k node of output layer, i=1, k=1;

Representing 3 BP network hidden layer activation functions respectively, all is the Sigmoidal function, and the Sigmoidal function expression is as follows:

$f_j^{\mathrm{\θ}}\left(u_j^{\mathrm{\θ}}\right)=\frac{1}{1+e^{{-u}_j^{\mathrm{\θ}}}}$

$f_j^{\mathrm{\γ}}\left(u_j^{\mathrm{\γ}}\right)=\frac{1}{1+e^{{-u}_j^{\mathrm{\γ}}}}$

Its derivative is:

$f_j^{\mathrm{\θ}\′}\left(u_j^{\mathrm{\θ}}\right)=\frac{1}{1+e^{{-u}_j^{\mathrm{\θ}}}}\frac{e^{{-u}_j^{\mathrm{\θ}}}}{1+e^{{-u}_j^{\mathrm{\θ}}}}=[1-f_j^{\mathrm{\θ}}\left(u_j^{\mathrm{\θ}}\right)]f_j^{\mathrm{\θ}}\left(u_j^{\mathrm{\θ}}\right)$

$f_j^{\mathrm{\γ}\′}\left(u_j^{\mathrm{\γ}}\right)=\frac{1}{1+e^{{-u}_j^{\mathrm{\γ}}}}\frac{e^{{-u}_j^{\mathrm{\γ}}}}{1+e^{{-u}_j^{\mathrm{\γ}}}}=[1-f_j^{\mathrm{\γ}}\left(u_j^{\mathrm{\γ}}\right)]f_j^{\mathrm{\γ}}\left(u_j^{\mathrm{\γ}}\right)$

Wherein,

Represent the input of 3 BP network hidden layer activation functions respectively, and

Be exponential function; Represent the output of 3 BP network hidden layer activation functions respectively,

Representing 3 BP network output layer activation functions respectively, all is the Sigmoidal function, and the Sigmoidal function expression is as follows:

$g^{\mathrm{\θ}}\left(u^{\′\mathrm{\θ}}\right)=\frac{1}{1+e^{{-u}^{\′\mathrm{\θ}}}}$

$g^{\mathrm{\γ}}\left(u^{\′\mathrm{\γ}}\right)=\frac{1}{1+e^{{-u}^{\′\mathrm{\γ}}}}$

Its derivative is:

$g^{\mathrm{\θ}\′}\left(u^{\′\mathrm{\θ}}\right)=\frac{1}{1+e^{{-u}^{\′\mathrm{\θ}}}}\frac{e^{{-u}^{\′\mathrm{\θ}}}}{1+e^{{-u}^{\′\mathrm{\θ}}}}=[1-g^{\mathrm{\θ}}\left(u^{\′\mathrm{\θ}}\right)]g^{\mathrm{\θ}}\left(u^{\′\mathrm{\θ}}\right)$

$g^{\mathrm{\γ}\′}\left(u^{\′\mathrm{\γ}}\right)=\frac{1}{1+e^{{-u}^{\′\mathrm{\γ}}}}\frac{e^{{-u}^{\′\mathrm{\γ}}}}{1+e^{{-u}^{\′\mathrm{\γ}}}}=[1-g^{\mathrm{\γ}}\left(u^{\′\mathrm{\γ}}\right)]g^{\mathrm{\γ}}\left(u^{\′\mathrm{\γ}}\right)$

Wherein, u ' ^θ, u ' ^γ,

Represent the input of 3 BP network output layer activation functions respectively, and $u^{\′\mathrm{\θ}}=\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\θ}}{f}_j^{\mathrm{\θ}}\left(w_{\mathrm{ij}}^{\mathrm{\θ}}{x}^{\mathrm{\θ}}\right),$ $u^{\′\mathrm{\γ}}=\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\γ}}{f}_j^{\mathrm{\γ}}\left(w_{\mathrm{ij}}^{\mathrm{\γ}}{x}^{\mathrm{\γ}}\right),$

Step 2: receive the gyro signal on X, Y, the Z axle respectively;

Step 3: judge whether to be in and enter the high overload environment, described high overload environment determination methods is: the accelerometer output according to strapdown inertial navigation system is judged, when accelerometer output is acceleration of gravity greater than setting threshold 5g and g, then be considered as entering the high overload environment, otherwise be considered as non-high overload environment;

Step 4: when judged result when entering the high overload environment, then 3 BP networks are operated in the simulation output state, simulation gyro signal output, and respectively with the output y of current up-to-date step 1 ^θ, y ^γ,

Calculate as the simulating signal output of gyro on X, Y, the Z axle and for navigation;

Step 5: when judged result is non-high overload environment, then 3 gyro outputs are used for navigation and calculate and the gyro under the non-high overload environment is exported as the online training sample of BP network, described online training sample comprises gyro output time correlation

With the gyro output under the non-high overload environment

Gyro accumulative total hits is p, makes p=p+1 again, if p is less than online number of training P, P=300 then returns step 2, if p equals 300, then utilize 300 training samples that the BP network is carried out online training, and with current online training result step of updating 1 described BP network connection weights

Return step 2 after the p zero clearing, described online training comprises the online training of X-axis, the online training of Y-axis and the online training of Z axle;

The step of the online training of described X-axis is as follows:

Step 5.1.1: make iterations n _θInitial value be 1, utilize random function Random () that the BP network hidden layer of X-axis gyro output under the simulation high overload environment and the weights of output layer are carried out initialization:

$w_{\mathrm{ij}}^{\mathrm{\θ}}=\begin{array}{cc}\mathrm{Random}(\·)& w_{\mathrm{ij}}^{\mathrm{\θ}}\∈\left[\mathrm{0,1}\right]\end{array}$

$w_{\mathrm{jk}}^{\mathrm{\θ}}=\begin{array}{cc}\mathrm{Random}(\·)& w_{\mathrm{jk}}^{\mathrm{\θ}}\∈\left[\mathrm{0,1}\right]\end{array}$

Step 5.1.2: calculate used 300 sample total errors in the current iteration process

$E_{A\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}{E}_{p\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=\frac{1}{2}\underset{p=1}{\overset{300}{\mathrm{\Σ}}}{(d_p^{\mathrm{\θ}}-y_{p\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}})}^2$

In the formula

Be the output of the BP network of X-axis gyro output under the simulation high overload environment when importing p sample in the current iteration process;

$y_{p\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=g^{\mathrm{\θ}}\left[\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\θ}}{f}_j^{\mathrm{\θ}}\left(w_{\mathrm{ij}}^{\mathrm{\θ}}{x}_p^{\mathrm{\θ}}\right)\right]$

Step 5.1.3: if Perhaps n _θ〉=10000, then adopt current

Under the simulation high overload environment of step of updating 1 in the BP network of X-axis gyro output

Otherwise enter step 5.1.4;

Step 5.1.4: make n _θ=n _θ+ 1, $w_{\mathrm{ij}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}+\mathrm{\Δ}{w}_{\mathrm{ij}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}},$ $w_{\mathrm{jk}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}+\mathrm{\Δ}{w}_{\mathrm{jk}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}$

Wherein

$\mathrm{\Δ}{w}_{\mathrm{ij}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=-0.58\frac{\∂E_A^{\mathrm{\θ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}$

$\mathrm{\Δ}{w}_{\mathrm{jk}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=-0.58\frac{\∂E_A^{\mathrm{\θ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}$

Wherein

$\frac{\∂E_A^{\mathrm{\θ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\θ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\underset{j=1}{\overset{10}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\θ}}}{\∂y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}\frac{\∂y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}{\∂u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}\frac{\∂u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}{\∂x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}\frac{\∂x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}{\∂u_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}\frac{\∂u_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}$

$=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\underset{j=1}{\overset{10}{\mathrm{\Σ}}}(d_p^{\mathrm{\θ}}-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}(1-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}{x}_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}(1-x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}})x_p^{\mathrm{\θ}}$

$\frac{\∂E_A^{\mathrm{\θ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\θ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\θ}}}{\∂y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}\frac{\∂u_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}{\∂u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}\frac{\∂u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}(d_p^{\mathrm{\θ}}-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})g^{\mathrm{\θ}\′}\left(u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}\right)x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}$

$=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}(d_p^{\mathrm{\θ}}-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}(1-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}$

In the formula

Represent the BP network output when p sample of input in the last iterative process,

Represent the input of the BP network output layer activation function when p sample of input in the last iterative process,

Represent the output of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent the input of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent that last iteration finishes the back hidden layer to the connection weights of output layer,

For last iteration finishes the back input layer to the connection weights of hidden layer, return step 5.1.2;

The step of the online training of described Y-axis is as follows:

Step 5.2.1: make iterations n _γInitial value be 1, utilize random function Random () that the BP network hidden layer of Y-axis gyro output under the simulation high overload environment and the weights of output layer are carried out initialization:

$w_{\mathrm{ij}}^{\mathrm{\γ}}=\begin{array}{cc}\mathrm{Random}(\·)& w_{\mathrm{ij}}^{\mathrm{\γ}}\∈\left[\mathrm{0,1}\right]\end{array}$

$w_{\mathrm{jk}}^{\mathrm{\γ}}=\begin{array}{cc}\mathrm{Random}(\·)& w_{\mathrm{jk}}^{\mathrm{\γ}}\∈\left[\mathrm{0,1}\right]\end{array}$

Step 5.2.2: calculate used 300 sample total errors in the current iteration process

$E_{A\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}{E}_{p\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=\frac{1}{2}\underset{p=1}{\overset{300}{\mathrm{\Σ}}}{(d_p^{\mathrm{\γ}}-y_{p\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}})}^2$

In the formula Be the output of the BP network of Y-axis gyro output under the simulation high overload environment when importing p sample in the current iteration process;

$y_{p\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=g^{\mathrm{\γ}}\left[\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\γ}}{f}_j^{\mathrm{\γ}}\left(w_{\mathrm{ij}}^{\mathrm{\γ}}{x}_p^{\mathrm{\γ}}\right)\right]$

Step 5.2.3: if

Perhaps n _γ〉=10000, then adopt current

Under the simulation high overload environment of step of updating 1 in the BP network of Y-axis gyro output

Otherwise enter step 5.2.4;

Step 5.2.4: make n _γ=n _γ+ 1, $w_{\mathrm{ij}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}+\mathrm{\Δ}{w}_{\mathrm{ij}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}},$ $w_{\mathrm{jk}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}+\mathrm{\Δ}{w}_{\mathrm{jk}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}$

Wherein

$\mathrm{\Δ}{w}_{\mathrm{ij}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=-0.58\frac{\∂E_A^{\mathrm{\γ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}$

$\mathrm{\Δ}{w}_{\mathrm{jk}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=-0.58\frac{\∂E_A^{\mathrm{\γ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}$

Wherein

$\frac{\∂E_A^{\mathrm{\γ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\γ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\underset{j=1}{\overset{10}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\γ}}}{\∂y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}\frac{\∂y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}{\∂u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}\frac{\∂u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}{\∂x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}\frac{\∂x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}{\∂u_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}\frac{\∂u_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}$

$=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\underset{j=1}{\overset{10}{\mathrm{\Σ}}}(d_p^{\mathrm{\γ}}-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}(1-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}{x}_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}(1-x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}})x_p^{\mathrm{\γ}}$

$\frac{\∂E_A^{\mathrm{\γ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\γ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\γ}}}{\∂y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}\frac{\∂y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}{\∂u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}\frac{\∂u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}(d_p^{\mathrm{\γ}}-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})g^{\mathrm{\γ}\′}\left(u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}\right)x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}$

$=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}(d_p^{\mathrm{\γ}}-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}(1-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}$

In the formula

Represent the BP network output when p sample of input in the last iterative process,

Represent the input of the BP network output layer activation function when p sample of input in the last iterative process,

Represent the output of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent the input of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent that last iteration finishes the back hidden layer to the connection weights of output layer,

For last iteration finishes the back input layer to the connection weights of hidden layer, return step 5.2.2;

The step of the online training of described Z axle is as follows:

Step 5.3.1: make iterations

Initial value be 1, utilize random function Random () that the BP network hidden layer of Z axle gyro output under the simulation high overload environment and the weights of output layer are carried out initialization:

Step 5.3.2: calculate used 300 sample total errors in the current iteration process

In the formula

Be the output of the BP network of Z axle gyro output under the simulation high overload environment when importing p sample in the current iteration process;

Step 5.3.3: if

Perhaps

Then adopt current

Under the simulation high overload environment of step of updating 1 in the BP network of Z axle gyro output

Otherwise enter step 5.3.4;

Step 5.3.4: order

Wherein

Wherein

In the formula

Represent the BP network output when p sample of input in the last iterative process,

Represent the input of the BP network output layer activation function when p sample of input in the last iterative process,

Represent the output of the BP network hidden layer activation function when p sample of input in the last iterative process, Represent the input of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent that last iteration finishes the back hidden layer to the connection weights of output layer,

For last iteration finishes the back input layer to the connection weights of hidden layer, return step 5.3.2.

Three gyros and three accelerometer quadratures are fixedly mounted in the package with navigational computer and other power supplys and control module and constitute strapdown inertial navigation system, strapdown inertial navigation system is directly installed on the carrier, the relative carrier mounting structure with accelerometer of the gyro of strapdown inertial navigation system as shown in Figure 1, strapdown inertial navigation system gyroscope signal intelligent filtering fundamental diagram is as shown in Figure 2.

Below narration is all at object peculiar to vessel, and namely carrier is general naval vessel.

1) the BP network structure of gyro signal artificial intelligence design:

Outfit form according to naval vessel strapdown inertial navigation system features of the object and gyro, three BP network structures of the present invention are the version of single input/list output and single hidden layer, the system that the is input as output time correlation of each BP network is output as gyro output information; According to the model analysis to the Ship Motion state that test of many times obtains, emulation is chosen 3000 groups of samples the BP network is carried out off-line training study, and the anticipation error of training is 0.0001 °, and maximum iteration time is 10000.Exporting the BP network with simulation X-axis gyro is example, extracts different training samples and carries out repeatedly l-G simulation test comparison of off-line, and part training result parameter sees Table 1,

Table 1

Table 1 is embodiment of the invention BP network training partial results statistical form, analyzes according to the training result statistical conditions, and hidden layer node is counted m and is defined as 10, and namely neural network is selected the structure of 1-10-1;

2) adopt the step 5 in the technical scheme to carry out the online training study of BP network parameter, because overload is uncertain constantly, the online training study process need of this part constantly repeats, to guarantee that the BP network parameter is up-to-date obtaining, the motion state that meets current naval vessel to greatest extent, thus the accuracy that gyro signal artificial intelligence module is exported improved.

The training time is about 0.078s in the simulation process, training result such as Fig. 6.Training utilizes the sampled data of 50s that the network that trains is tested after finishing, test result such as Fig. 7, dotted line is BP network analog output among the figure, solid line is the actual output of system, BP network output is fluctuateed in small range with the difference of desired output as seen from the figure, shows that the BP network can effectively simulate naval vessel actual motion state.

3) gyro signal artificial intelligence module

Increase a strapdown inertial navigation system gyroscope signal intelligent filtering dedicated thread in the computing machine of naval vessel strapdown inertial navigation system, the main thread parallel running is calculated in this thread and strapdown inertial navigation system navigation; Under non-overload environment, 5 carry out the online training study of the BP network parameter of gyro signal artificial intelligence set by step, and judge the current motion state on naval vessel by the output valve of sense acceleration meter; When the naval vessel is under the high overload environment, control gyro signal intelligent filter thread switches to the simulation output state, navigation is calculated main thread and is adopted the output of gyro signal artificial intelligence module to replace the actual output of gyro, and naval vessel gyro signal artificial intelligence module application schematic diagram as shown in Figure 8.

4) test result analysis

It is the gyro output information of 45s that l-G simulation test adopts time span, and the high overload motion state appears in carrier when 41.5s, utilizes the output of gyro signal artificial intelligence functional module output alternative system this moment.Be output as example with simulation X-axis gyro, the actual output of system is as Fig. 9, under the high overload environment, utilize figure as a result such as Figure 10 of the output of gyro signal artificial intelligence functional module output alternative system, Figure 11 is real system output and the auxiliary output of gyro signal artificial intelligence functional module result contrast, the solid line representative result that actual output is resolved according to gyro, dotted line represents the auxiliary output of gyro signal artificial intelligence functional module result.The laboratory is the semi-physical simulation result show, utilize the BP network can simulate strapdown inertial navigation system output under the high sea situation environment of high frequency, overload impact, thus avoid because impact, the gyro output noise causes under the vibration condition strapdown inertial navigation system resolution error.

Claims (1)

1. the strapdown inertial navigation system gyroscope signal intelligent filtering method under the high overload environment is characterized by:

Step 1: set up 3 BP network structures, described 3 BP network initial configurations comprise the BP network of Z axle gyro output under the BP network of X-axis gyro output under the simulation high overload environment, the BP network of simulating Y-axis gyro output under the high overload environment and the simulation high overload environment, 3 BP network initial configurations are: the three-layer network version of single input/single output and single hidden layer, the hidden layer node number is m, m=10; The input of 3 BP networks is respectively the gyro output time correlation x on X, Y, the Z axle ^θ, x ^γ,

Output is respectively the gyro output y on X, Y, the Z axle ^θ, y ^γ,

Wherein x, y represent input, the output of BP network respectively, subscript θ, γ,

Represent inputing or outputing of corresponding X, Y, Z axle respectively; Relation between the output of 3 BP networks and the input is respectively:

$y^{\mathrm{\θ}}=g^{\mathrm{\θ}}\left[\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\θ}}{f}_j^{\mathrm{\θ}}\left(w_{\mathrm{ij}}^{\mathrm{\θ}}{x}^{\mathrm{\θ}}\right)\right]$

$y^{\mathrm{\γ}}=g^{\mathrm{\γ}}\left[\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\γ}}{f}_j^{\mathrm{\γ}}\left(w_{\mathrm{ij}}^{\mathrm{\γ}}{x}^{\mathrm{\γ}}\right)\right]$

In the formula

Represent 3 BP network input layers respectively to the connection weights of hidden layer, subscript i, j represent i node of input layer and j node of hidden layer respectively;

Represent 3 BP network hidden layers respectively to the connection weights between the output layer, subscript k represents k node of output layer, i=1, k=1;

Representing 3 BP network hidden layer activation functions respectively, all is the Sigmoidal function, and the Sigmoidal function expression is as follows:

$f_j^{\mathrm{\θ}}\left(u_j^{\mathrm{\θ}}\right)=\frac{1}{1+e^{{-u}_j^{\mathrm{\θ}}}}$

$f_j^{\mathrm{\γ}}\left(u_j^{\mathrm{\γ}}\right)=\frac{1}{1+e^{{-u}_j^{\mathrm{\γ}}}}$

Its derivative is:

$f_j^{\mathrm{\θ}\′}\left(u_j^{\mathrm{\θ}}\right)=\frac{1}{1+e^{{-u}_j^{\mathrm{\θ}}}}\frac{e^{{-u}_j^{\mathrm{\θ}}}}{1+e^{{-u}_j^{\mathrm{\θ}}}}=[1-f_j^{\mathrm{\θ}}\left(u_j^{\mathrm{\θ}}\right)]f_j^{\mathrm{\θ}}\left(u_j^{\mathrm{\θ}}\right)$

$f_j^{\mathrm{\γ}\′}\left(u_j^{\mathrm{\γ}}\right)=\frac{1}{1+e^{{-u}_j^{\mathrm{\γ}}}}\frac{e^{{-u}_j^{\mathrm{\γ}}}}{1+e^{{-u}_j^{\mathrm{\γ}}}}=[1-f_j^{\mathrm{\γ}}\left(u_j^{\mathrm{\γ}}\right)]f_j^{\mathrm{\γ}}\left(u_j^{\mathrm{\γ}}\right)$

Wherein,

Represent the input of 3 BP network hidden layer activation functions respectively, and

Be exponential function;

Represent the output of 3 BP network hidden layer activation functions respectively

Representing 3 BP network output layer activation functions respectively, all is the Sigmoidal function, and the Sigmoidal function expression is as follows:

$g^{\mathrm{\θ}}\left(u^{\′\mathrm{\θ}}\right)=\frac{1}{1+e^{{-u}^{\′\mathrm{\θ}}}}$

$g^{\mathrm{\γ}}\left(u^{\′\mathrm{\γ}}\right)=\frac{1}{1+e^{{-u}^{\′\mathrm{\γ}}}}$

Its derivative is:

$g^{\mathrm{\θ}\′}\left(u^{\′\mathrm{\θ}}\right)=\frac{1}{1+e^{{-u}^{\′\mathrm{\θ}}}}\frac{e^{{-u}^{\′\mathrm{\θ}}}}{1+e^{{-u}^{\′\mathrm{\θ}}}}=[1-g^{\mathrm{\θ}}\left(u^{\′\mathrm{\θ}}\right)]g^{\mathrm{\θ}}\left(u^{\′\mathrm{\θ}}\right)$

$g^{\mathrm{\γ}\′}\left(u^{\′\mathrm{\γ}}\right)=\frac{1}{1+e^{{-u}^{\′\mathrm{\γ}}}}\frac{e^{{-u}^{\′\mathrm{\γ}}}}{1+e^{{-u}^{\′\mathrm{\γ}}}}=[1-g^{\mathrm{\γ}}\left(u^{\′\mathrm{\γ}}\right)]g^{\mathrm{\γ}}\left(u^{\′\mathrm{\γ}}\right)$

Wherein, u ^'θ, u ^{' γ},

Represent the input of 3 BP network output layer activation functions respectively, and $u^{\′\mathrm{\θ}}=\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\θ}}{f}_j^{\mathrm{\θ}}\left(w_{\mathrm{ij}}^{\mathrm{\θ}}{x}^{\mathrm{\θ}}\right),$ $u^{\′\mathrm{\γ}}=\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\γ}}{f}_j^{\mathrm{\γ}}\left(w_{\mathrm{ij}}^{\mathrm{\γ}}{x}^{\mathrm{\γ}}\right),$

Step 2: receive the gyro signal on X, Y, the Z axle respectively;

Step 3: judge whether to be in and enter the high overload environment, described high overload environment determination methods is: the accelerometer output according to strapdown inertial navigation system is judged, when accelerometer output is acceleration of gravity greater than setting threshold 5g and g, then be considered as entering the high overload environment, otherwise be considered as non-high overload environment;

Step 4: when judged result when entering the high overload environment, then 3 BP networks are operated in the simulation output state, simulation gyro signal output, and respectively with the output y of current up-to-date step 1 ^θ, y ^γ,

Calculate as the simulating signal output of gyro on X, Y, the Z axle and for navigation;

Step 5: when judged result is non-high overload environment, then 3 gyro outputs are used for navigation and calculate and the gyro under the non-high overload environment is exported as the online training sample of BP network, described online training sample comprises gyro output time correlation

With the gyro output under the non-high overload environment

Gyro accumulative total hits is p, makes p=p+1 again, if p is less than online number of training P, P=300 then returns step 2, if p equals 300, then utilize 300 training samples that the BP network is carried out online training, and with current online training result step of updating 1 described BP network connection weights

Return step 2 after the p zero clearing, described online training comprises the online training of X-axis, the online training of Y-axis and the online training of Z axle;

The step of the online training of described X-axis is as follows:

Step 5.1.1: make iterations n _θInitial value be 1, utilize random function Random () that the BP network hidden layer of X-axis gyro output under the simulation high overload environment and the weights of output layer are carried out initialization:

$w_{\mathrm{ij}}^{\mathrm{\θ}}=\begin{array}{cc}\mathrm{Random}(\·)& w_{\mathrm{ij}}^{\mathrm{\θ}}\∈\left[\mathrm{0,1}\right]\end{array}$

$w_{\mathrm{jk}}^{\mathrm{\θ}}=\begin{array}{cc}\mathrm{Random}(\·)& w_{\mathrm{jk}}^{\mathrm{\θ}}\∈\left[\mathrm{0,1}\right]\end{array}$

Step 5.1.2: calculate used 300 sample total errors in the current iteration process

$E_{A\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}{E}_{p\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=\frac{1}{2}\underset{p=1}{\overset{300}{\mathrm{\Σ}}}{(d_p^{\mathrm{\θ}}-y_{p\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}})}^2$

In the formula

Exceed for when importing p sample, simulating in the current iteration process

$y_{p\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=g^{\mathrm{\θ}}\left[\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\θ}}{f}_j^{\mathrm{\θ}}\left(w_{\mathrm{ij}}^{\mathrm{\θ}}{x}_p^{\mathrm{\θ}}\right)\right]$

Step 5.1.3: if

Perhaps n _θ〉=10000, then adopt current

Under the simulation high overload environment of step of updating 1 in the BP network of X-axis gyro output

Otherwise enter step 5.1.4;

Step 5.1.4: make n _θ=n _θ+ 1, $w_{\mathrm{ij}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}+\mathrm{\Δ}{w}_{\mathrm{ij}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}},$ $w_{\mathrm{jk}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}+\mathrm{\Δ}{w}_{\mathrm{jk}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}$

Wherein

$\mathrm{\Δ}{w}_{\mathrm{ij}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=-0.58\frac{\∂E_A^{\mathrm{\θ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}$

$\mathrm{\Δ}{w}_{\mathrm{jk}\left(n_{\mathrm{\θ}}\right)}^{\mathrm{\θ}}=-0.58\frac{\∂E_A^{\mathrm{\θ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}$

Wherein

$\frac{\∂E_A^{\mathrm{\θ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\θ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\underset{j=1}{\overset{10}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\θ}}}{\∂y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}\frac{\∂y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}{\∂u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}\frac{\∂u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}{\∂x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}\frac{\∂x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}{\∂u_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}\frac{\∂u_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}$

$=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\underset{j=1}{\overset{10}{\mathrm{\Σ}}}(d_p^{\mathrm{\θ}}-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}(1-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}{x}_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}(1-x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}})x_p^{\mathrm{\θ}}$

$\frac{\∂E_A^{\mathrm{\θ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\θ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\θ}}}{\∂y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}\frac{\∂y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}{\∂u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}\frac{\∂u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}}=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}(d_p^{\mathrm{\θ}}-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})g^{\mathrm{\θ}\′}\left(u_{p(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}\right)x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}$

$=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}(d_p^{\mathrm{\θ}}-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}}(1-y_{p(n_{\mathrm{\θ}}-1)}^{\mathrm{\θ}})x_{\mathrm{jp}(n_{\mathrm{\θ}}-1)}^{\′\mathrm{\θ}}$

In the formula

Represent the BP network output when p sample of input in the last iterative process,

Represent the input of the BP network output layer activation function when p sample of input in the last iterative process,

Represent the output of the BP network hidden layer activation function when p sample of input in the last iterative process, Represent the input of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent that last iteration finishes the back hidden layer to the connection weights of output layer,

For last iteration finishes the back input layer to the connection weights of hidden layer, return step 5.1.2;

The step of the online training of described Y-axis is as follows:

Step 5.2.1: make iterations n _γInitial value be 1, utilize random function Random () that the BP network hidden layer of Y-axis gyro output under the simulation high overload environment and the weights of output layer are carried out initialization:

$w_{\mathrm{ij}}^{\mathrm{\γ}}=\begin{array}{c}\begin{array}{cc}\mathrm{Random}(\·)& w_{\mathrm{ij}}^{\mathrm{\γ}}\∈\left[\mathrm{0,1}\right]\end{array}\end{array}$

$w_{\mathrm{jk}}^{\mathrm{\γ}}=\begin{array}{c}\begin{array}{cc}\mathrm{Random}(\·)& w_{\mathrm{jk}}^{\mathrm{\γ}}\∈\left[\mathrm{0,1}\right]\end{array}\end{array}$

Step 5.2.2: calculate used 300 sample total errors in the current iteration process

$E_{A\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}{E}_{p\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=\frac{1}{2}\underset{p=1}{\overset{300}{\mathrm{\Σ}}}{(d_p^{\mathrm{\γ}}-y_{p\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}})}^2$

In the formula

Be the output of the BP network of Y-axis gyro output under the simulation high overload environment when importing p sample in the current iteration process;

$y_{p\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=g^{\mathrm{\γ}}\left[\underset{j=1}{\overset{10}{\mathrm{\Σ}}}{w}_{\mathrm{jk}}^{\mathrm{\γ}}{f}_j^{\mathrm{\γ}}\left(w_{\mathrm{ij}}^{\mathrm{\γ}}{x}_p^{\mathrm{\γ}}\right)\right]$

Step 5.2.3: if

Perhaps n _γ〉=10000, then adopt current

Under the simulation high overload environment of step of updating 1 in the BP network of Y-axis gyro output

Otherwise enter step 5.2.4;

Step 5.2.4: make n _γ=n _γ+ 1, $w_{\mathrm{ij}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)+}^{\mathrm{\γ}}\mathrm{\Δ}{w}_{\mathrm{ij}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}},$ $w_{\mathrm{jk}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}+\mathrm{\Δ}{w}_{\mathrm{jk}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}$

Wherein

$\mathrm{\Δ}{w}_{\mathrm{ij}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=-0.58\frac{\∂E_A^{\mathrm{\γ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}$

$\mathrm{\Δ}{w}_{\mathrm{jk}\left(n_{\mathrm{\γ}}\right)}^{\mathrm{\γ}}=-0.58\frac{\∂E_A^{\mathrm{\γ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}$

Wherein

$\frac{\∂E_A^{\mathrm{\γ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\γ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\underset{j=1}{\overset{10}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\γ}}}{\∂y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}\frac{\∂y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}{\∂u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}\frac{\∂u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}{\∂x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}\frac{\∂x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}{\∂u_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}\frac{\∂u_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}{\∂w_{\mathrm{ij}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}$

$=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\underset{j=1}{\overset{10}{\mathrm{\Σ}}}(d_p^{\mathrm{\γ}}-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}(1-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}{x}_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}(1-x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}})x_p^{\mathrm{\γ}}$

$\frac{\∂E_A^{\mathrm{\γ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\γ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=\underset{p=1}{\overset{300}{\mathrm{\Σ}}}\frac{\∂E_p^{\mathrm{\γ}}}{\∂y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}\frac{\∂y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}{\∂u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}\frac{\∂u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}}{\∂w_{\mathrm{jk}(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}}=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}(d_p^{\mathrm{\γ}}-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})g^{\mathrm{\γ}\′}\left(u_{p(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}\right)x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}$

$=-\underset{p=1}{\overset{300}{\mathrm{\Σ}}}(d_p^{\mathrm{\γ}}-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}}(1-y_{p(n_{\mathrm{\γ}}-1)}^{\mathrm{\γ}})x_{\mathrm{jp}(n_{\mathrm{\γ}}-1)}^{\′\mathrm{\γ}}$

In the formula

Represent the BP network output when p sample of input in the last iterative process, Represent the input of the BP network output layer activation function when p sample of input in the last iterative process,

Represent the output of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent the input of the BP network hidden layer activation function when p sample of input in the last iterative process, Represent that last iteration finishes the back hidden layer to the connection weights of output layer, For last iteration finishes the back input layer to the connection weights of hidden layer, return step 5.2.2;

The step of the online training of described Z axle is as follows:

Step 5.3.1: make iterations

Initial value be 1, utilize random function Random () that the BP network hidden layer of Z axle gyro output under the simulation high overload environment and the weights of output layer are carried out initialization:

Step 5.3.2: calculate used 300 sample total errors in the current iteration process

In the formula

Be the output of the BP network of Z axle gyro output under the simulation high overload environment when importing p sample in the current iteration process;

Step 5.3.3: if Perhaps Then adopt current

Under the simulation high overload environment of step of updating 1 in the BP network of Z axle gyro output Otherwise enter step 5.3.4;

Step 5.3.4: order

Wherein

Wherein

In the formula Represent the BP network output when p sample of input in the last iterative process,

Represent the input of the BP network output layer activation function when p sample of input in the last iterative process,

Represent the output of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent the input of the BP network hidden layer activation function when p sample of input in the last iterative process,

Represent that last iteration finishes the back hidden layer to the connection weights of output layer,

For last iteration finishes the back input layer to the connection weights of hidden layer, return step 5.3.2.

Images