2. Previous and this work
โข SINS/DVL
โข INS/GPS
โข SINS/GPS/DVL
โข SINS/TAN/DVL/MCP
implemented in this paper
3. Why NN with KF
โข Kalman filter diverge from
estimate and can not deal
colured noise underwater
โข Neural Network are good
with dealing non linearity
(main point is always inertial
measurement with external
global fixes)
4. Configuration of system
โข SINS has no substansive
plateform so acceleration
and angular velocity are
there pose, position and
velocity are calculated at
100Hz
โข Dvl provid velocity relative to
sea bottom
โข heading is found by Dvl
โข position fixes are by TAN
5. Filter equation - Mathematical model
โข east-north-vertical coordinate
and state vector is
vilocity errors attitude
angle errors
lattitude
longtude and
high errors
accelerometer
biases
gyro drifts
real
postion
position
obtrained
from TAN
SINS
TAN
6. BP neural network (intro)
โข Main functions are
adapation, generation and
powerful fault tolerance
โข Neural network uses non-
linearity and differential
function to train weights.
7. BP NN algorithm
โข input(x) and output(y) with
connection wij and wjk
โข Training process:
w and b small random values,
determine actual output(for x, y),
lastly weights are adjusted to
minimize errors, gives us result
followed by gradient descent of
cost function
iterate until the cost function smaller
the e(set value)
8. USE OF BP NN in this paper
โข Recall phase
correction by NN(sample with
enough percision) are added to
Kalman filter
input of NN is highly error
producing sensors
โข Observation and prediction
vector as one input
300 samples to train network offline
9. Simulations
โข Matlab 6.5 and VC++6.0
tool is used
โข flat plane under certain
depth with some coditions
โข linear velocity is 4kn,
heading 45, longitude and
latitude 165 and 32, drift
5/h and noise 10/h,
random constant bias
50ug and error 50ug, dvl
covariance error 0.5m/s,
compass 3 and TAN 50m
10. conclusion
โข In this approach, the errors
in
โข the classical federated
Kalman filter estimation
are corrected by
โข the BP neural network
which is trained off line.
AUV position
โข error is substantially
reduced and the precision
of the
โข underwater navigation is
greatly improved.