INTRODUCTION
Liquid rocket engines experiment is a extremely sophisticated integrated technology
and huge systematic project. Compared with other large scientific experiments,
it has particularly high comprehensiveness encompassing large numbers of subjects
such as rocket engine expertise, monitoring and control technology, propellant
chemistry technology, cryogenic technology, vacuum technology, highaltitude
environment simulation technology, environmental monitoring technology, environmental
technology and organization and management science (Figueroa
and Schmalzel, 2006). Because of complex structure, high precision and hefty
expense, experiment failure can lead to huge losses and even equipment damage.
According to the fault records form Aerospace Testing Technology Institute,
the failure rate due to experiment system fault was 20%. The early detection
of faults can help avoid failure from spreading, reduce system shutdown and
prevent accident involving human fatality and material damage (Wu,
2005). Therefore, it has great economic and security significance for early
fault detection and timely preventive maintenance of the liquid rocket engine
testbed.
Our liquid rocket engines testbed still adopts off line monitoring method
in the present in China which merely forces on several main parameters, while
may result in fallacious prediction even erroneous failure detection relating
to the sensors fault (Wu, 2005). Many widely employed
methods based on statistics are simple and reliable, however, whose accuracy
rely heavily on statistical results of testbed. And it is difficult to establish
accurate mathematical model by mathematical analysis on account of nonlinear
and instability of the liquid rocket engine testbed. In recent years, as computer
technology and artificial intelligence constantly developed, many new theories
and methods in the field of fault detection and diagnosis are applied to propulsion
system. Fuzzy hypersphere neural network based realtime monitoring failure
system is presented (Huang et al., 1999) which
is verified that it outperforms the BP neural network in the article. Neural
network based realtime fault detection algorithm in ground test process is
implemented with MATLAB (Huang et al., 2007)
which performs effectiveness in engine failure detection through many offline
evaluation and realtime online test. A Support Vector Machines (SVM) based
multifault classifier is established for data mining of liquid rocket engine
steady state test (Han and Hu, 2007) which is verified
by experiment that the algorithm has excellent classification and antijamming
ability and requires few training data.
To deal with the defects of difficultly determined neural network structure and easily immerging in partial minimum frequently, a novel adaptive particle swarm optimization relevance vector machine (APSORVM) is proposed which is based on liquid rocket engines testbed fault detection method. Through, mapping the raw data to high dimension space with kernel function, it solves nonlinear in the high dimension space to avoid the lower linear inseparable problem. It not only resolves over learning problems in virtue of small samples and nonlinear but also generates better generalization ability by solving sparse model.
The main contributions in this study are the development of the following:
• 
A novel RVM kernel parameter selection method is proposed,
reflecting faster computing speeds and locating the global optimum 
• 
A novel liquid rocket engines testbed fault detection method was suggested,
reducing the alarm time and promoting the system reliability 
THE RELEVANCE VECTOR MACHINE
RVM, like SVM, has a sparse probabilistic model (Tipping,
2001). However, RVM appeases to be advantageous against SVM due to Bayesian
treatment which does not suffer from the basic limitations of SVM. The RVM trains
in the Bayesian framework, obtains the priori probability of weight by a set
of hyperparameters and find the optimal value via iterative algorithm. RVM
applied Bayesian inference based on Gaussian process methods to SVM, derives
probability distribution and makes kernel function free from Mercer. It has
good performance (Tipping and Lawrence, 2005) on function
regression and classification.
For regression, given a set of input vector
and relevant output vector ,
supervised learning aims at designing a model with those training data and prior
knowledge (Tipping and Faul, 2003). Based on new input
vector x_{n}, the model can forecasted relevant output vector y(x_{n}).
t_{n} are observed output which can be taken as a unknown function y(x,
w) contained the Gaussian noise with σ^{2} variance:
where, ε is independently distributed noise; w is adjustable parameters weight. So, obtain:
where, Φ_{i}(x) = K(x, x_{i}). The selection of kernel function will be free from Mercer. Thus, we can choose the most popular kernels such as Gaussian kernel, polynomial kernel, Radial Basis Function (RBF) kernel, etc. The likelihood function corresponding training set is :
Where:
and:
that:
The way got the optimal weight w of Empirical Risk Minimization (ERM) can result
in over learned. To avoid this situation, given a priori conditional probability
distribution using sparse Bayesian method for the weight w:
Meanwhile, for hyperparameters α and noise variance β ≡ σ^{2} definite hyperprior distribution as Γ distribution:
Where:
In normal circumstances, parameter a, b, c, d is very tiny, it can provide
a = b = c = d = 0. So, it will get consistent hyperprior.
Then, according to the Bayes rule, it can find posteriori formula for all of unknown parameters as following:
Given to new observed point x_{*}, the distribution of corresponding target forecasted value t_{*} is:
Considering:
Thus:
where, covariance matrix is Σ = (σ^{2} Φ^{T}Φ+A)^{1}, A = diag(α_{0}, α_{1},..., α_{N}), mean vector is μ = σ^{2}ΣΦ^{T}t.
This study use the peak of delta function to approximate the hyperprior p(α,
σ^{2}t) in t he above formula. Aim at forecasting, we do not care
p(α, σ^{2})≈δ(α_{MP}, σ^{2}_{MP}),
where α, σ^{2} possible values are α_{MP}, σ^{2}_{MP}.
Focus on:
Thus, RVM problem is converted into posterior mode for hyperparameters problem, that is, seeking the maximum of posteriori hyperparameters which is equal to seeking the maximum of α and β. In the condition of consistent hyperprior, it only needs to take the maximum of p(tα, σ^{2}). So, obtain:
To obtain p(tα, σ^{2}), it need to integrate out α and β with following iterative formula:
where, γ_{i} = 1α_{i}N_{ii}, N_{ii} are the ith diagonal elements of posteriori weight covariance matrix.
Ideally, if given an new input x_{*}, it can predict target through following steps:
Corresponding to the new observations, the predicted output of RVM is y(x_{*}, μ).
SELECTION OF RVM KERNEL PARAMETER
Yet it has never been proposed any analytic instructional method for RVM kernel
parameter selection (Li et al., 2010; Tao
et al., 2008). At present, empirical method and ergodic searching
have been mostly used to study kernel parameter setting. Ergodic searching has
certain blindness for parameters setting and demands enormous sacrifices of
time.
Characteristics analysis of RVM kernel parameter: In this study, based on standard function sinc, effects of kernel parameters on training results is discussed.

Fig. 1: 
Regression performance diagram 
Table 1: 
Regression performance of different σ 

The function of sinc is:
where, noise is random noise in the range [1, 1]. Even taken 100 sample points
as samples, this paper adopted 5fold crossvalidation to calculate training
error and test error. The Gaussian kernel used for the construction of the basis
matrices is K(x, y) = exp[xy/σ^{2}].
From the Table 1, it can conclude that training error approximates
zero as σ approaches zero which means all samples can be compact fitted.
Whereas, we must notice that corresponding test error approaches infinity, the
number of corresponding relevance vector is 100, that is to say, all the training
samples are relevance vector, for test samples are 100. At this stage there
is over learned and RVM lost learning ability. All above arguments show ERM
based on traditional learning approach such as neural network cannot ensure
excellent generalization ability. The sizes of relevance vector get smaller
with the increase of σ (in a small range) and training error increasing
but test error decreasing. That means the generalization ability of RVM is improving.
When σ approximates 4.2, it can get minimum test error. Figure
1 illustrates minimum test error is not coincide with minimum training error
which verifies traditional ERM cannot guarantee excellent generalization ability.
When σ reach a certain value (nearly 6 in the Table 1),
the sizes of relevance vector and corresponding training error and test error
are increasing again. That indicates the performance of discrimination and generalizations
gets worse. Actually, the generalizations ability of RVM based on Gaussian kernel
goes from low to high and to low.
Adaptive particle swarm optimization: After RVM kernel parameter characteristics
analysis, a novel method about optimizing RVM kernel parameter with adaptive
particle swarm optimization is proposed. Particle Swarm Optimization (PSO) based
on swarm intelligence theory is an optimization technique for locating the global
optimum (Tu et al., 2011) which produces swarm
intelligence to guide optimization with cooperation and competition among particles.
This algorithm simulates birds foraging behavior, considers each individual
always maintain the optimal distance with adjacent ones during activity. It
promotes the coevolution in the virtue of information sharing. Comparing with
other intelligent optimization algorithms, PSO has an excellent performance
on parallel searching and meanwhile it has few parameters and shape convergence
speed. In PSO, each optimization problem is regarded as searching a flight particle
in the space. Flight direction and distance of particle depend on velocity and
optimal objective function determines fitness. Particles dynamically adjust
their speed based on flying experience and find the optimal solution by iteration.
According to the above problem, defined PSO fitness function as:
where, w_{i} is relevance vector weight, y_{i} is truth value,
is predicted value.
The implementing steps of this optimization process are:
• 
Step 1: V_{max} = 1, V_{min} = 1,
c_{1} = c_{2} = 1.5, iteration step is 100, population size
is 20. Definite mutation operator as 0.9 on the basic PSO to prevent from
immerging in partial minimum on account of iterative efficiency reduced
in the post course of the optimization procedure 
• 
Step 2: Random generate velocity and position of each particle 
• 
Step 3: Calculate Fitness according to fitness function 
• 
Step 4: Compare the current fitness value and the best
fitness value in history for each particle. If the current fitness value
is smaller, then take it as the best experienced individual fitness value 
• 
Step 5: Calculate the minimum of the best fitness value of all
particle experience and record it as the global best fitness value 
• 
Step 6: Update particle velocity and position and limit each new
velocity and position 
• 
Step 7: Random generate mutation factor. If reach the variation
conditions, reinitialize the particles 
• 
Step 8: If reach the maximum of iteration step, stop searching
and return the best fitness value and the position. Else turn to step 3 
Table 2: 
Optimization results of APSO 

Following the above steps, 10 optimization results are give in Table
2.
From Table 2, the experimental results show that σ makes a nearly perfect fit with the optimal value at a high probability, meanwhile which prove consistent relevance vector number and close sparse level (7/100). From what has been discussed above, it may safely draw the conclusion that the proposed APSO is efficiently applicable to the problem of RVM kernel parameter optimization.
PROJECT EXAMPLE
Establishing mathematical model is the emphasis and difficulty of complex and
nonlinear dynamic systems fault detection (Yu et al.,
2009). For a project of liquid rocket engines testbed, hydrogen supply
system is the key part of rocket engine testbed, whose work status will have
a directly impact on the engine test. For security, it have to realtime monitor
many significant parameters, such as oxygen valve outlet pressure (Pejy), oxygen
tank pressure (Pxy), prepump pressure of oxygen (Pohy), etc. which ensure reliability
of entire system. Because system conditions rapidly changed before ignition
start, the value of parameters mentioned above changed abruptly. It is necessary
to use sensors realtime monitoring signal change, in order to prevent unnecessary
damage.

Fig. 3: 
The results of training Pejy via APSORVM 
Table 3: 
Pejy training parameters and results 

Here, it demonstrates how the proposed APSORVM framework is efficiently applicable to the project of liquid rocket engines testbed fault detection. This paper focus our analysis on the parameter of Pejy and model it using 37 times normal test data which are truncated 10 seconds data in precooling phase before ignition with 50 Hz sampling frequency. The total sample points are 500. The modeling process is as follows:
• 
Data preparation. First of all, make data preprocessing and
normalize to [0, 1]. Besides, select Gaussian kernel and optimize kernel
parameter using APSORVM. Return optimal kernel parameter σ 
• 
Modeling using the theory of section 2. Simplified process is as Fig.
2 
• 
Calculating prediction error and evaluating upper and lower limit of model 
• 
According to Chebyshev inequality P(xμ≥nσ)≤1/n^{2},
via APSORVM model, obtain: 
In practice, making estimated standard deviation S instead of the standard
deviation σ. The threshold range of sensor signal characteristics can be
formulate as:
where, w^{T} is relevance vector, φ(x_{i}) relevance vector kernel matrix, n = 3. The probability of falling into this interval is 99.74%. The results of training Pejy model via APSORVM are shown in Fig. 3. Training parameters and results are presented in Table 3.
The experimental results indicated that APSORVM has excellent modeling ability (fitness is 8/500), while reduces detection time.
Furthermore, for verifying the performance of model presented via APSORVM, it employed on the actual data of an oxygen valve leakage. Using the method proposed in this literature, alarm occurred at 17.53 sec which is 0.95 sec ahead of alarm time via red threshold detection method.
CONCLUSION
In this study, a novel APSORVM kernel parameter method is proposed for augmenting RVM kernel parameter selection by adaptive PSO. What this paper done make up for amending the defects in the analytic instructional method for RVM kernel parameter selection. Results show that kernel parameter will converge in a tiny region using analytic instructional method proposed instead of experience and ergodic searching. It is efficiently applicable to the project of liquid rocket engines testbed fault detection and can rapidly detect failure which meet the engineering requirements of realtime and reliability with high practical significance. According to the different observer points, it can flexibly analysis data features and select the appropriate parameters. It will have extensively effect.
As further study, the proposed model remains to integrate consider the pressure and flow direction among different observer points.