For the sake of implementing effectively detection and rapid exact identification for the chemical agents in sea-battlefield, the automatic detection technique using the neural network (NN) information fusion for the chemical agents, is studied.
By Zhang Minghu,
Yang Hongyu, Bai Xuelian and Chen
Hongmin
Dept. of
Underwater Weaponry and Chemical Defense, Dalian Naval Academy, Dalian, China
The Detachment
of Warship Training, Dalian Naval Academy, Dalian, China
Abstract:
For the sake
of implementing effectively detection and rapid exact identification for the
chemical agents in sea-battlefield, the automatic detection technique using the
neural network (NN) information fusion for the chemical agents, is studied.
Firstly, the information fusion treatment model and system frame are analyzed;
Secondly, the modeling means and chemical agents feature extraction are
studied, connecting the wavelet analysis with the NN organically, and based on
the wavelet transfer and NN, the system of the speedy features extraction and
identification for chemical agents, the NN distinguishing chemical agent
system, is founded; Thirdly, the hardware realization and software component
are introduced in brief; Lastly, the method is workable, and of the high
identification accuracy, remarkable generalization capability, good stability,
fast speed, and high reliability. And the method has the general military
application value.
Introduction:
The conception of information fusion began in
1970s’ initial stages, rooted in the requirement of the C3 I (Command, Control,
Communication and Intelligence) system in military domain, then instituted the
multi-sources correlation or multi-sensors mix data fusion, and the
technologies were established in 1980s. The JDL(Joint Directors of
Laboratories) of the DOD(department of defense) of the US defined information
fusion as such process from military application: those data and information
from multi-sensors and multi-info-sources add association, correlation and
combination, to obtain the accurate position estimation, identity estimation,
and the proper integrity estimation for the battlefield instance and menace and
its importance degree. Waltz and Llinas put up supplement and modification the
above definition: the information fusion is a sort of multilayer and/or multi-sided
treatment process, the process is detection, association, correlation,
estimation and combination to obtain the accurate state estimation, identity
recognition, the whole situation estimation and menace estimation.
Treatment
model:
The
information fusion treatment model is for a set of the treatment processes’
description, which is the constitution of the system function units, not
refered the physics frame and software realization of each of the units, and
the treatment processes also allow feedback, viz regulate the fusion frames
based on the decision-making. Fig.1 is the information fusion treatment model
that was put forward by the data fusion work group in the US, which it plays
important influence to understand the basic conception of the information
fusion.
The
information fusion system frame:
The research
of the system frame of the information fusion includes two parts: hiberarchy
and architecture. The hiberarchy mostly based on the information to analyze the
fusion system; and the architecture mostly based on the hardware.
Image Attribute: Information Fusion Tool / Source: Lockheed Martin
Hiberarchy:
The
information fusion system may be compartmentalized according to administrative
levels, but there are more opinions. At present, the universal acceptant
opinion is three layers fusion structure, namely data layer, feature layer and
decision-making layer. The hiberarchy of the information fusion is
compartmentalized according to the information abstract degree. In the practice
engineering application of the multi-sensors information fusion (MSIF) system,
the hiberarchy model, which was established and adopted, should synthetically
consider the sensors performance, system calculation ability, communications
bandwidth, expectation nicety rate, and existence fund power. Moreover, the
establishment of the hiberarchy based on the information can better confirm the
architecture based on the system hardware.
Architecture:
The system
hardware architecture are approximately divided to three species: concentration
style, distribution style and admixture style. The architecture adopted is
entirely the various practice requirements. When the architecture is designed,
it’s confirmed by the established system hiberarchy, at the same time, the many
supporting technologies must also be considered, such as data traffic, database
management, human-computer port, and sensor management.
Military
application model construction
Neural network
(NN) is a new technology for pattern recognition, which can distinguish
nonlinear complex objects using self-adapting mode. It has advantages of high
legitimacy and strong anti-interference capacity. At presents, there are many
applied studies on the BP NN and RBF NN in the chemical agents and environment
monitoring, the BP is the most comprehensive application in the aspect of
spectrum discrimination and is the most successful technology, and appears many
improved BP algorithm. And the RBF NN has become the research hots-pot at present,
which has higher discrimination capability and faster learning speed. Kohonen
network and adaptive resonance NN are belonging to self-organizing and
self-adapting NN, which can cultivate independent learning ability, and be
applied in the adaptability training of the equipment to new surroundings. NN
have high recognition capability and mature theory, and the algorithm is
complex, only combined with the method of the efficient feature extraction or
selection, can we get ahead in the practical application.
Modeling
means:
Takes example
for the naval ships chemical detection (NSCD), for the sake of implementing
effectively detection and rapid exact identification for the chemical agents in
sea-battlefield, the automatic detection technique -NN information fusion
technology- for the chemical agents, is studied. The model of the NN
information fusion system is built. At the same time, connecting the wavelet
analysis with the NN organically, and based on the wavelet transfer and NN, the
system of the speedy features extraction and identification for chemical
agents, the neural network distinguishing chemical agent (NNDCA) system, is
founded. In order to establish the NSCD system of high speed decision-making,
advanced principle, higher precision and sensitivity, automation, and
intelligence, and heightening the NSCD technology.
Selecting the
sample data of the chemical agents to carry on the feature extraction, and this
is the premise of the selecting recognition parameter and the structuring
distinguishing NN. Carries on the wavelet energy spectrum analysis for the
chemical agents sample parameter, to obtain the corresponding wavelet energy
spectrum characteristic vector of the parameter residual error data, which are
available in structuring the study sample of the recognition network.
Founded on
certain quantity training samples collection (usually is the symptom, chemical
agent of data set), to carry on the training for the NN to obtain the
expectation recognition network, then to comply with the current recognition
input, the chemical agent is distinguished for the system. Before study and recognition,
it usually does justice to the primary distinguishing data and the training sample
data. The goal is to provide the appropriate recognition input for the
recognition network and the training sample, and the usable feature vector for
the NN distinguishing. When the chemical agent enters the chemical sensors, the
complex film in the sensors adsorb the chemical agent molecule. According to
the differences of the chemical agent compound, it may cause the resonance
frequencies change through the mini-sensors. The mini-sensors have the
different frequency feature response signal for the different chemical agent.
This kind of response is recorded and transmitted to the NN, which has trained
and joins the chemical sensors, to examine the hairlike distinction in these
signals. The chemical agent feature is recorded in the database with
comparison, distinction and confirmation the chemical agent type.
The NNDCA for
the NSCD system based on the MSIF is to extract or select the feature of the
chemical agent in sea battlefield, according as the feature of the every
chemical agent which is measured by the multi-sensors in the system, and make
use of the NN’ ability of the non-linear mapping and pattern distinguishing,
eventually distinguish the chemical agent to be measured by synthesis and
analysis. Before using the NN to distinguish the chemical agents, the NN is
trained to anticipant recognition network by some quantitative training samples
sets.
The NNDCA was
built up by the numerical value calculating, NN’s learning and training. The
input network’s node-number is the dimension of the input characteristic
vectors; the output’s is amount of the chemical agents; the middle-tier’s
should be selected by the request of the training sample-set size and training
error. The precondition of the distinguishing preferences and constructing
recognition NN is the characteristic extracting for these samples data, the
eigenvector of the normalized wavelet energy distribution can be obtained by
the analyzing wavelet energy distribution for the data of the chemical agent
sample parameters, and which may be regarded as the calculation basis of the
parameter recognition confidence-degree or samples comprehensive discrete degree.
The corresponding wavelet energy distribution eigenvector of the parameter
residual data can be obtained by treating the chemical agent sample residual
data, which can construct the learning sample of the recognition NN. If the
training effect is imperfect, it need readjust the parameters of the NN,
repartition the training samples, and retrain the NN, till attain to the
satisfying effect.
Chemical
agents feature extraction:
In the process
of distinguishing the chemical agents, we can get the practical data of many
parameters by chemical sensors. These data contain rich information of the
chemical agents. Feature extraction or selection is the powerful tool to reduce
the mode dimension. Reducing dimension of the primitive features can help
minimize the wrong recognition rate of the classified device. Via wavelet
package analysis technology, we can record the signals energy data of various
frequencies, thus we get the frequency energy feature, which distinguishes one
from another. Choosing feature parameter is based on choosing the sample data’s
features. With parameter features’ comprehensive discrete degree or recognition
credit degree as a criterion, we can get a group of better feature parameters
to recognize. We have known that the different chemical agents obtain the
different frequency energy feature vectors. They reflect the different chemical
agents, and which provides the practical feature vectors for the recognition of
chemical agents based on the NN. According to the feature vector to work out
the sort decision-making, i.e., to confirm the chemical agent that it is which kind
of chemical agents.
Hardware
realization:
In order to
realize the NNDCA, to heighten the stability and reliability of the NSCD
system, and to adapt to the requirement of miniaturization, the calculating
method of the NNDCA should be transplanted to the chip of the hardware, and get
the practical application. For the interface-control part of the NNDCA and NSCD
system, the embedded designing thought may be used, the various single-chip or
embedded computers can be adopted as the running platform. For the system of
the biggish data quantity, it can adopt the DSP device to realize the beginning
pretreatment function, and the embedded computer to realize the data traffic
and interface management of the sub assemblies in the NNDCA classifiers and NSCD
system. In order to make the system run speedily and obtain the finer real-time
ability, the NN classifiers can also adopt the realization project of the
digital circuit based on FPGA.
Software
component:
The software exploitation of the NNDCA is the
quite complex software engineering. The software is mostly composed of the many
modules, such as the chemical agent recognition process evaluation module,
NNDCA module, data processing module, data management module, maintenance
module, and database. The module of the NNDCA and the database are the cores of
the wholly formed software.
Conclusions
Using the
information fusion and wavelet signal means to distinguish the chemical agent,
which the conclusions are as follows:
(1) The
multi-dimensional information of the chemical agents can be obtained by using
the measurement system based on the multi-sensor fusion; the data from the
multi-sensor can be fused by the neutral networks, the results show that the
method can remarkably fall the effects of the factors, interference,
concentration, and condition, etc, for the measurement results, and heighten
the accuracy and credibility of the measurement results.
(2) The
anticipative recognition of NN can be gained by training, testing, and verifying, on
the basis of the relatively processing for the learning sample data and its
feature extraction or selection, it is completely feasible that the analyses
for the chemical agents with the NNDCA system based on the MSIF technology.
Acknowledgements:
This work was financially supported in part by
the national defense Basic Research Foundation of China (435B956), the National
Defense Pre-Research Found of China(41101050403), and the 3rd Academic
Pre-Research Funds of Dalian Naval Academy of the 2110 Project
(DLJY-XY2015004).
References
[1] G. Zhang.
Chemical sensors, No. 1(2011), p. 54.
[2] Q. Pang,
X. Yu, Y. Cheng, et al. Acta Automatica Sinica, No.4 (2003), p. 599.
[3] M. Zhang,
G. Pang, H. Liu, et al. Journal of Guangxi Normal University: Natural Science
Edition,
No.1 (2008),
p. 220.
[4] G. Wang,
Zh. Zhang. Electrical Measurement & Instrumentation, No. 2(2006), p. 1.
[5] H. Zhou.
Journal of Chifeng University (Natural Science Edition), No. 5(2011), p. 96.
[6] F. Wang,
Sh. Huang, Zh. Han. Aeronautical Computing Technique, No. 1(2009), p. 102.
[7] Y. Ao, B.
Wang. Machine Tool & Hydraulics, No. 13(2014), p. 168.
[8] J. Liu, P.
Wang, L. Yang. Journal of Beijing University of Technology, No. 6(2014), p.
843.
[9] X. Tian,
Zh. Ma, Sh. Tan. Journal of Sichuan Ordnance, No. 5(2015), p. 90.
[10] X. Jian,
H. Jia, L. Shi. Chinese Journal of Construction Machinery, No. 2(2009), p. 227.
[11] H. Ge, L.
Tian. Application Research of Computers, No. 1(2009), p. 19.
[12] M. Liu,
L. Peng. Microelectronics & Computers, No. 1(2009), p. 166.
© Atlantis
Press. This article is distributed under the terms of the
Creative Commons
Attribution License / Download the Pdf - LINK
