Fault Diagnosis Method of Spacecraft Measurement and Control
Equipment Based on Artificial Intelligence Technology
Hongjian Guo
12
, Chun Qi
1
, Yanguo Chen
3
, Jun Wei
2
and Haipeng Liu
2
1
Xi'an Jiaotong University, Xi'an, China
2
Xi’an Satellite Control Center, Xi'an, China
3
Troops 94587, Lianyungang, China
Keywords: Artificial intelligence; Aerospace monitoring; Fault diagnosis.
Abstract: To solve the problems of fault diagnosis methods such as low versatility, difficulty in selecting feature
parameters, and low efficiency of reasoning in spacecraft equipment, a fault diagnosis method for space
equipment based on artificial intelligence is proposed. Feature selection and extraction, and then use the
support vector machine to optimize the selected parameters, and finally through the application of
simulation in the fault diagnosis of multiple types of space measurement and control equipment to verify the
effectiveness of the system.
1 INTRODUCTION
The aerospace monitoring and control equipment
has a complex structure, cross-coupling of
subsystems, and a wide range of failure modes.
Research on reliability modeling and analysis
methods for complex aerospace monitoring and
control equipment is an important guarantee for
improving the automation level of aerospace
monitoring and control equipment operation and
management[1-3]. For complex large-scale
measurement and control equipment, relying only on
equipment operators using traditional methods for
fault diagnosis and troubleshooting has greater
limitations, applying artificial intelligence
technology to the field of aerospace monitoring and
control equipment fault diagnosis, through the
establishment of fault diagnosis expert system, the
equipment Faults are quickly and accurately located,
and expert-level maintenance recommendations are
proposed to provide new methods for fault diagnosis
of measurement and control equipment.
The paper proposed a general method for fault
diagnosis of spacecraft monitoring and control
equipment based on genetic programming and
support vector machines. From the selection of the
original parameters of fault features of genetic
programming, the optimization of characteristic
parameters, SVM for fault diagnosis and other steps,
the fault diagnosis method for space equipment
based on artificial intelligence is studied. , To
provide an effective guarantee for the safe and
reliable operation of aerospace monitoring and
control equipment.
2 GENETIC PLANNING
ALGORITHM
2.1 Endpoint Sets and Function Node
Sets
The set of endpoints contains the different types of
variables and constants provided to the GP system
and is the input to the program generated by the GP
algorithm. The set of endpoints is the most basic
element of the problem environment and results. For
different problems, the meaning of the elements is
also different. The selection of a function node set is
a statement, an operation, and a function suitable for
a GP operation, including an operator, a function
operation, and some expressions.
2.2 Initial Population and Common
Generation Methods
The initial population of the GP algorithm consists
of a number of randomly generated individuals.
These individuals consist of a given function and
various possible symbolic expressions. The process
of generating the initial population can be seen as a
blind search process in the program space. When
generating the initial group of individuals, a function
is first randomly selected in the function node set as
the root node of the syntax tree; then, the same
number of child nodes are selected according to the
number of independent variables handled by the
function. For each subtree starting from the child
node, an element can be randomly selected from the
union of the function node set and the end point set
as the node of the subtree. If the selected function is
a function, it will be repeated. The above operation
process; if the selected end point, the subtree stops
growing.
2.3 Fitness Evaluation
When the individuals in the population replicate,
crossover, and mutate, the evaluation scale of the
individual in the GP algorithm is called a fitness
function. The fitness evaluation of the GP algorithm
uses standardized fitness, initial fitness, adjusted
fitness, and normalized fitness.
Standardization fitness is consistent with the
maximization of fitness in genetic algorithms, and
can be expressed in the following simple form:
)()(
max
irriS =
1
In the formula, S(i) is the output value calculated
by the individual i fitness, rmax is the maximum
original fitness, and r(i) is the original fitness of
individual i.
Primitive fitness is a measure of the natural
description of the problem, usually obtained by
directly calculating the absolute error between the
individual's output and the expected output,
=
=
M
j
jCjiSir
1
)(),()(
2
In the formula, S(i,j) is the calculated output of
the individual i at the jth input value; C(j) is the
target expected value corresponding to the jth input
value; M is the number of training samples.
The standardized fitness is adjusted, and the
adjusted fitness a(i) of the individual i is calculated
by the following equation.
)(1
1
)(
iS
ia
+
=
3
In general, S(i)0, then a(i)[0,1]. Therefore,
the greater the adjusted fitness value, the better the
individual. Adjusting the degree of fitness is better
than the standard level of fitness for the best
individual, especially when the standard fitness
approaches zero, adjusting the degree of fitness can
amplify small differences in standard fitness.
The normalized fitness degree is a selection
method based on the fitness proportion, which is
calculated by adjusting the fitness degree. The
concrete expression is
=
=
M
k
ka
ia
in
1
)(
)(
)(
4
In the formula, M is the size of the population.
The normalized fitness degree has the following
three ideal characteristics: n(i)[0,1]; the greater
the fitness value, the better the individual is:
=
M
k
ka
1
)(
=1.
2.4 Selecting Strategies
The GP algorithm selection strategy includes fitness
proportion selection method, fitness ranking
selection method, roulette selection method and
tournament selection method.
2.5 Genetic Manipulation
GP algorithm genetic operations include: copy,
crossover, and mutation. Among them, the crossover
operation is based on the rule that the higher the
fitness value is, the better the probability of being
selected is. From the current population, two parent
individuals are randomly selected. The tree structure
of two parent individuals is shown in Fig.1; then,
from two A node is randomly selected as a cross
point in the tree, and the entire sub-tree below the
cross point is taken as a cross section (as shown by a
dashed box in Fig.1 ).
Figure 1: Tree structure of two parent individuals.
Figure 2: Tree structure of two individuals after crossover.
Two cross-sections selected in the two parent
individuals are exchanged to generate two trees. The
new tree is the two children of the next generation.
The tree structure of the two individuals after the
crossover operation is shown in Fig. 2.
2.6 Termination Criteria
The GP program is a process of highly parallel, local
control and decentralized processing. The state of
each stage is only determined by the genetic
program group. The fitness index drives the group to
constantly adapt to the changes of the environment,
making the whole group adaptable development of.
3 AVIATION MEASUREMENT
AND CONTROL EQUIPMENT
FAULT FEATURE SELECTION
AND EXTRACTION
3.1 The Selection of Fault
Characteristics of the Original
Parameters
In order to weaken the influence of factors such as
process parameters and operating conditions of the
aerospace measurement and control equipment on
the diagnostic effect, and at the same time ensure
sufficient sensitivity to equipment defects and faults,
the selection of genetic operating parameters is
usually selected to be sensitive to fault information.
Given that dimensionless indices are used as genetic
operands, genetic planning requires that the original
indicators be optimized and combined. Therefore, it
is necessary to prove that the combined feature
parameters are still dimensionless parameters [4-5].
There are 2 dimensionless indicators, in the
following format:
)(
)(
12
11
1
da
da
p =
5
)(
)(
22
21
2
db
db
p =
6
In the formulas,
1
a
and
2
a
represent two variables
with dimension d1; b1 and b2 represent two
variables with dimension d2. Let
21
ppV +=
, then
)(
)(
)()(
)()()()(
)(
)(
)(
)(
2122
2112
2212
21122211
22
21
12
11
21
ddX
ddX
dbda
dbdadbda
db
db
da
da
ppV =
+
=+=+=
7
From the formula (7), we can see that
V is the
ratio of the two variables X
12
and X
22
with the same
dimension (all
d
1
d
2
), so it is still a dimensionless
index. By the same token, it can be shown that after
the arithmetic operations are performed on the
dimensionless indicators, the results are still
dimensionless.
3.2 Determination of Fitness Function
For the n-type d-dimensional sample set, contains N
samples x1, x2, ..., xN, where N1 belongs to class
ω
1, denoted as
Ξ
1; N2 belongs to class
ω
2,
denoted as
Ξ
2; Nn belongs to class
ω
n, denoted as
Ξ
n .
The average value of mi for mi samples is:
Ξ
=
x
i
x
N
m
1
, i=1,2,...,n (8)
The intra-class dispersion Di and the total intra-
class average dispersion Dw are:
T
ii
x
i
mxmxD ))(( =
Ξ
i=1,2,...,n (9)
=
=
n
i
iw
D
n
D
1
1
(10)
Sample class dispersion Dbij is
T
jij
x
ibij
mmmmD ))(( =
Ξ
ij (11)
Fitness function
ω
D
D
F
bij
)min(
=
(12)
Where Dbij represents the class spacing between
the i-th and j-th classes. The numerator represents
the minimum value of the dispersion between
classes, and the denominator represents the average
value of the dispersion within the class.
4 SUPPORT VECTOR MACHINE
PARAMETER OPTIMIZATION
Select 50 sets of data from the data samples of A1
and A2 as the training samples of the support vector
machine. Set the initial population number N of the
genetic algorithm to 100, the crossover probability
Pc=0.7, the mutation probability Pm=0.15, and
terminate the algebra For 400, training and
parameter optimization are performed. The
optimization results are: C3=47.26, σ3=1.74.
5 SIMULATION
In the experimental process, the collected data is
first calculated according to the calculation formula
of each characteristic parameter of the aerospace
measurement and control equipment, and margin
index L, peak factor C, kurtosis factor K, waveform
of each group of measurement and control
equipment can be obtained. RMS ratio Rv and other
parameters, Table 1 gives the calculation results of
10 domain parameter values under different working
conditions.
Table 1: Time-frequency parameter values when the
aviation measurement and control equipment is damaged.
Paramete
r
Number
Rv C K L
1 78.4652 1.9878 2.4908 2.2325
2 97.4409 1.9080 2.5642 2.9842
3 96.7694 1.9631 2.9067 2.9487
4 89.0543 1.9884 2.4554 2.8225
5 98.6545 1.8774 2.3211 2.3221
6 87.4321 1.8998 3.2323 2.9885
7 103.7689 1.9990 3.1221 3.3417
8 101.2321 1.9567 2.9003 2.5663
9 100.4531 1.9899 2.3425 2.7888
1
103.2321 1.9996 2.7888 3.4175
In the experiment process, the data from the
lubricating oil spectral analysis data of the aerospace
measurement and control equipment in the complete
working phase, and according to the time sequence
of the spectrum analysis to obtain the original data
sequence of the metal content of the aviation
measurement and control equipment, as shown in
Fig. 3.
Figure 3: Raw data sequence of metal content.
The data sequence is processed relative to each
other, and then the data is subjected to zero
processing to obtain a new Al metal relative content
time series {xi}. Relative data processing, and then
the data after the zero treatment to get the new
relative content of Al time series. According to the
calculation formula of the embedding dimension, the
FPE value corresponding to different embedding
dimensions shown in Fig. 4 can be calculated. From
Figure 4, we can see that when it is equal to 10, the
FPE value is the smallest, that is, the best embedding
dimension is 10.
Figure 4: The FPE values corresponding to different
embedding dimensions.
6 CONCLUSIONS
Genetic programming is used to select the features
of aerospace monitoring and control equipment, and
the parameters of aerospace monitoring and control
equipment are optimized using regression SVM,
which makes the optimized state diagnosis model of
aerospace monitoring and control equipment have
higher accuracy. Under the given device parameter
alert value, the support vector machine can be used
to diagnose the status of aerospace monitoring and
control equipment in a relatively long range, which
can provide an important basis for the health
monitoring of space monitoring and control
equipment.
REFERENCES
1. WANG Wanyu,TAO Sunjie,FENG Xuxiang. Design
of fault diagnosis expert system for remote sensing
satellite receiving system [ J]. Telecommunication
Engineering,2015,55(5):491-496.
2. ZHOU Qijun,ZHAO Qiuying,ZHU Mingming. TT&C
systemfault diagnosis method based on fault tree[J].
Modern Electronics Technique,2015,38(7):103-106.
3. WANG Weihong,LI Jun. Improved edit distance
algorithm base on local variability[J]. Computer
Engineering,2015,41(7):294-298.
4. WANG Hua,LIU Huanmin,DUAN Huifen. Case-
basedreasoning method in fault diagnosis expert
system of inertial navigation system [ J]. Journal of
Chinese Inertial Technology,2009,17(5):614-617.
5. Baker J.E.. Adaptive selection methods for genetic
algorithms[J], Proc. of the Intel Conf. on Genetic
Algorithms and Their Applications, 2008:101~111