Download as pdf or txt
Download as pdf or txt
You are on page 1of 10

2016 IEEE Second International Conference on Big Data Computing Service and Applications

Fault-Diagnosis for Reciprocating Compressors


Using Big Data

Keerqinhu∗ , Guanqiu Qi† , Wei-Tek Tsai†‡ , Yi Hong∗ , Wenxiang Wang∗ , Guangxin Hou∗ , Zhiqin Zhu§
∗ Research Center of China National Offshore Oil Corporation, Beijing, 100010, China
† Schoolof Computing, Informatics, and Decision Systems Engineering
Arizona State University, Tempe, AZ, 85287, USA
‡ State Key Laboratory of Software Development Environment

School of Computer Science and Engineering, Beihang University, Beijing, 100191, China
§ College of Automation, Chongqing University, Chongqing, 400044, China

keerqh@cnooc.com.cn, guanqiuq@asu.edu, tsai7@yahoo.com, {826422484, 326260390}@qq.com,


hougx@cnooc.com.cn, zhiqinzhu@126.com

Abstract—Reciprocating compressors are widely used in the rig, where it ignited and exploded, engulfing the platform [23].
petroleum industry, and a small fault in reciprocating com- The explosion killed 11 workers and injured 16 others. The
pressors may cause serious issues in operation. Monitoring and sinking of platform resulted in a massive offshore oil spill in
detecting potential faults help compressors to continue normal the Gulf of Mexico. Thus, real-time fault diagnosis system is
operation. This paper proposes a fault-diagnosis system for com- used to reduce the maintenance costs and improve working
pressors using machine-learning techniques to detect potential
efficiency of equipment [13], [16], [18].
faults. The system has been evaluated using 100TB operation
data collected from China National Offshore Oil Corporation, and Until now, a large amount of methods based on pressure,
the data are first de-noised, coded, and then SVM classification vibration, and acoustic emission (AE) signals have been used
is applied, with 50% of data used for training, the remaining for to diagnose faults in reciprocating compressors [22]. Its in-
testing. The results demonstrated that the system can efficiently
stallation condition and working condition vary. Therefore,
diagnose potential faults in compressors with 80% accuracy.
it is difficult to map between fault and working status on
Keywords—Reciprocating Compressor, Big Data, RPCA, SVM; reciprocating compressors. With the development of online
testing technology [14], [20], [26], real-time information needs
to be processed by hundreds of engineers in offshore oil fields.
I. I NTRODUCTION So the analysis costs are expensive.
Reciprocating compressors are widely used in the In most cases, fault diagnosis of reciprocating compressors
petroleum industry. It is important to keep reciprocating com- uses a real-time signal-processing model [12], [4]. Recipro-
pressor working properly. A reciprocating compressor operates cating compressors in offshore oil platforms usually work in
at a high-temperature, high-pressure, flammable, explosive, a salty and foggy environment, and the harsh environment
corrosive working environment, and it is used in the offshore makes traditional solutions not accurate. In addition, the real-
oil and gas production. So comparing to other equipment used time information often contains different parameters, such as
on land, reciprocating compressor requires a high reliability vibration, temperature, and displacement. The relationships
[22], [28], [27]. among these parameters are complex as they are related to
each other. Thus, it is difficult to determine the root cause
Currently monitoring of reciprocating compressors is done
for a system failure. It is important to identify the relevant
by human inspection. Regular maintenance and fault diagnosis
information from the large number of data from heterogeneous
after a system failure are common methods. Thus, it is difficult
multiple sources.
for current methods to predict the potential faults in advance.
Traditional fault-diagnosis methods focus on monitoring,
The US oil and gas industry spent about 30% of the total
and they can detect faults only after the compressor fails
repair cost ($75 billion) in 1980 due to improper equipment
to work. This paper proposes a fault-diagnosis system using
maintenance [11]. Furthermore, it is important to identify and
the big-data approach. An automatic recognition model is
repair any failed equipment. Many tiny faults that are not fixed
proposed based on a data-driven classification method. The
in time finally result in disasters. For instance, the explosion of
main contributions of this paper are as follows:
British North Sea Piper Alpha platform shocked the world on
July 6, 1988 [24]. The disaster killed 167 out of 228 workers. 1) A new fault-classification model for compressors based
The total insured loss was about £1.7 billion (US$3.4 billion). on machine learning;
The explosion was caused by the failure of a frozen pump was 2) Multiple category SVM recognizes normal and fault data
not identified and fixed in time. The pump was down at the to identify potential faults;
running time, and eventually caused an irreparable disaster. 3) Experimentation with data collected from real systems.
The Deepwater Horizon drilling rig exploded in the Gulf of
Mexico on April 20, 2010. High-pressure methane gas from the This paper is structured as follows: Section II presents the
well expanded into the drilling riser and rose into the drilling proposed framework; Section III discusses the fault-diagnosis

978-1-5090-2251-9/16 $31.00 © 2016 IEEE 72


DOI 10.1109/BigDataService.2016.27
process; Section IV evaluates the proposed framework and The high-level load balancer differentiates and allocates data
analyzes experiment results; and Section V concludes this analysis tasks to different clusters based on domain informa-
paper. tion. The compressors from the same ocean areas have the
same or similar conditions. To improve data analysis, it is
II. A RCHITECTURE better to cluster the compressor data based on the ocean area.
The specialized data are stored in the same or closely related
China’s coastline along the Pacific Ocean is 14,500 kilome- databases. The clustered data are assigned to one or several
tres (9,000 mi) long, and is bounded by the Bohai, Yellow, East closely related servers to process. The servers need to use the
and South China Seas. China has a large number of oil and gas designated databases corresponding to the clustered data. And
reserves under the bottom of the sea. Many drilling platforms the same type of data from the one domain is assigned to one or
are built to explore oil and gas. As four seas are vast and several clusters that specialize in analyzing this type of data.
span from north to south, the condition of diverse ocean area In a low-level, each cluster has its own local load balancer
varies a lot from surface water to deep oceanic trenches. So and it will dispatch data analysis tasks to different servers
compressors in different ocean areas face different conditions. within the cluster. The analyzed data of each server is saved
A large number of data needs to be processed to monitor in local database and shared to other servers in same cluster.
the status of compressors as each compressor generates about Data collector merges all analyzed data from each cluster. The
3 GB of data per hour. A drilling platform often has hundreds finalized results will be sent to technicians.
of compressors, but it is often located far away from the land
Concurrent algorithm shown in Algorithm I is proposed
and does not have enough computing capacity to process the
to solve the distribution and collection of analysis workloads.
data. Thus large volumes of data are sent from platforms to a
Based on domain information, unclassified data from machine
cloud for processing.
mi is assigned to domain di . Then domain di assigns its
Two types of data can be obtained from compressors: workloads to cluster ci , according to the computation capacity
structured and unstructured data. Structured data are related of each cluster and data type. In each cluster, the assigned data
to the status of compressors such as temperature, speed, and are analyzed. The results from each cluster are merged to form
acceleration; unstructured data are from video surveillance. the final results.
This paper focuses on analyzing structured data. The proposed
data analysis system has two parts: Algorithm 1 Concurrent Algorithm for Data Analysis
Input:
1) Learning part: It analyzes data to develop a model to
Unclassified data from machine mi , domain di , cluster ci
predict the future working status;
Output:
2) Analysis part: It uses the model to generated to predict
the status of compressors and to identify potential faults. n result ri , intermediate analyzed result iri
Analyzed
1: ci = i=1 mi
The proposed analysis framework is shown in Figure 1. 2: for analyze all ci do
The data analysis process runs in a cloud environment. The 3: return iri
proposed framework has three parts: 4: end for
n
5: di = i=1 ci
• Data Management: It saves all the data from com- 6: ri = (n, di , iri )
pressors in repositories hosted on a cloud. 7: for merge all di do
• Data Analysis: It uses data mining and machine 8: return ri
9: end for
learning methods to filter, classify, and analyze data.
First, it filters all noises in the original data. Then
the filtered data are classified and analyzed. Finally, it
formalizes a monitoring model based on the analyzed III. FAULT D IAGNOSIS
data. The formalized model is used to predict potential
faults. As more data are collected, the model will be Fault affects normal operation of industrial system. Al-
continuously updated. though varieties of approaches are applied to fault diagnosis,
accidents still happen in industrial production. Due to the
• Visualization: The analyzed data are presented as complexity of industrial system, fault diagnosis is always a
figures and tables to technicians for decision making, challenge work. Especially for reciprocating compressor in
real-time support, and machine management. petroleum industry, any small fault may cause serious disaster.
Numerous irreparable disasters of petroleum industry hap-
Figure 2 shows the concurrent design for reciprocating
pened in history. To avoid these disasters, different approaches
compressor data analysis. This is similar to the scalability
of fault diagnosis are applied to reciprocating compressor.
architecture commonly used in SaaS [15], [21]. Many ma-
Existing methods focus on monitoring, and cannot detect
chines work at the same time and the corresponding data are
potential faults at the first moment, due to the large size of
continuously sent to a cloud for analysis. The data will be
data. To efficiently diagnose faults of reciprocating compressor,
classified and assigned to different clusters for analysis. Each
a novel big-data based framework is proposed to analyze the
cluster has multiple servers to handle different tasks in parallel.
large-size data of reciprocating compressor. The proposed big-
The two-level architecture not only automatically balances data based framework is designed to recognize fault state
the workloads across multiple clusters and servers, but also and classify identified faults of reciprocating compressor for
scales up with increasing loads with automated expansion. repairing.

73

Fig. 1. Big Data Analysis Architecture in a Cloud Environment

A. Fault Diagnosis sparse coefficient get from the training data are used for SVM
classifier training. And the sparse coefficients get from the
Figure 3 shows the fault-diagnosis process. The operating
testing data can classified by the trained classifier. As each
data are acquired by sensors of reciprocating compressor.
sparse coefficient is corresponding to a source data vector,
The acquired data are about the vibration status of different
when the sparse coefficient is classified, the data are classified.
component, temperature, and humidity. The source data of
reciprocating compressor D consists of a few data vectors
d1 , d2 , ...dn , where di is the operating data at one moment. B. Robust Principal Component Analysis
The source data has been randomly divided into training data The size of source data obtained from server is large. And
and testing data [17], [19]. In this paper, half of data are used the source data have many noises. Directly doing fault classi-
for training and the rest of data are used for testing. Then both fication on source data with noises may not produce accurate
training and testing data are de-noised by RPCA [25], [5]. All results. A common strategy is to de-noise the source data first.
spike noises of data are isolated by RPCA. Then an online Principal Component Analysis (PCA) is a way to de-noise
dictionary learning and sparse coding process are implemented data using Gaussian distributed noise, widely used in signal
for data feature extraction. All source data are used in online and image processing [1]. However, signals of compressors
dictionary learning [6], [8]. The calculated sparse coefficients usually come with spike noise, that is difficult to be de-noised
of training data are used in multiple category SVM classifier by PCA. To solve spike noise issues, a novel RPCA method
training [9], [2]. The trained SVM classifier is used in testing is proposed for processing compressor signals.
data classification.
1) Problem Formulization of Data De-noise: In compressor
The fault-diagnosis processes of reciprocating compressor data, the rank of data matrix is low in most cases. In that case,
data shown in Figure 3 has three main processes. the data from compressor can be decomposed into a low rank
• Robust Principal Component Analysis (RPCA): As matrix and a sparse spike noise matrix. Then the source data
an unsupervised dimensionality reduction method, it is can be re-written as follows:
applied to source data to reduce spike noise. D=L+S , (1)
• Dictionary Learning: It extracts the features of the
where D is the source data of compressor, L is the low rank
de-noised operating data.
matrix of signal, and S is a sparse noise matrix. To find L, and
• SVM Classification: It is an efficient classifier for S as the best description of source data, a restriction function
vector classification with high accuracy. The leaned is proposed in Equation 2.
data are classified by a multiple category based SVM
classifier. min rank(L) + γS0 , (2)
L,S

All the data are firstly de-noised by RPCA and used for If one solves the problem in Equation 2 for appropriate γ, one
training the dictionary. The trained dictionary is used for may recover the pair (L, S) that generated the data D. However,
sparse representation of all the de-noised data. At last, the Equation 2 is non-convex, minimize both rank(L) and S0

74
       

    

      

   

               


               
 

                           

               

   

Fig. 2. Two-level Data Analysis Framework


Fig. 3. Fault Diagnosis of Reciprocating Compressor Data

are NP-hard to approximate. To solve this non-convex problem, In ALM for RPCA, we define
Peng and Marelaxes L0-norm to L1-norm and replace rank D = (L, S)
with a nuclear norm:
f (x) = A ∗ +λE1 , (4)
min L∗ + λS1 ,
L,S (3) h(x) = D − S − L
This relaxation can be motivated by observing that L∗ + Then the Lagrange function is show in Equation 5
λS1 is the convex envelope of rank(L) + λS0 . So that,
the L0 and S0 can be calculated by L0 , S0 = arg min L∗ + L(L, S, Y, μ) = L ∗ +λS1 + < Y, D − L − S >
L,S , (5)
λS1 . where D = L0 +S0 , L0 is the low rank matrix and S0 +μ/2D − L − SF
is a sparse matrix.
The Optimization Flow is just like the general ALM
2) Augmented Lagrange Multiplier (ALM) for RPCA: To method. The initialization Y = Y0 ∗ is to make the objective
solve that problem, ALM algorithm [7] is proposed for RPCA. function value< D, Y0 ∗ > reasonably large.

75
According to the objective Equation 5, the objective func- Then the problem can be written as follows:
tion of S can be rewritten as: m
 2
f (S) = λ||S||1 + < Y, D − L − S > min li − Dicxi 2 + λxi 1 , (11)
2 , (6) i=1
+μ/2 · ||D − L − S||F
Optimize 11 can find a xi for each li .

f(S) = λ||S||1 + < Y, D − L − S > The optimization algorithm contains two steps. The first
2 , (7) step is sparse coding process, shown in Algorithm III. And the
+μ/2 · ||D − L − S||F + (μ/2) ||μ−1 Y ||2 second step is dictionary learning process, shown in Algorithm
IV.

f(S) = λ||S||1 + (μ/2) (2(μ−1 Y · (D − L − S)) Algorithm 3 Online Dictionary Learning


2 , (8)
+||D − L − S||F + ||μ−1 Y ||2 ) Input:
x ∈ Rm ∼ p(x)
Output:
2 A trained dictionary Dic
f(S) = (λ/μ) ||S|| 1 + 1/2||S − (D − L − μ−1 Y )||F , (9)
1: A0 ∈ Rk×k ,B0 ∈ Rk×k
2: for t = 1 to T do do
Then the RPCA problem can be solved by a itera- 3: Sparse coding: compute using LARS
tive algorithm. For Algorithm II, any accumulation point Δ 2
(L∗,S∗)(Lk ∗,Sk ∗) is an optimal solution to the RPCA problem xt = arg min lt − Dict xt 2 + λxt 1
xt ∈Rk
and the convergence rate is at least When the low rank matrix 4: At ← At−1 + αt αtT
5: Bt ← Bt−1 + αt αtT
Algorithm 2 RPCA via ALM 6: Compute Dict using Algorithm IV using Dict−1 as
Input: warm restart, so that
Source data D ∈ Rm×n Dt = arg min 1t ( 12 T r(DT DAt ) − T r(DT Bt ))
Output: D
end for
(L∗k+1 and S∗k+1 )
7:

1: Y0∗ = sgn(D)/J(sgn(D));
2: while not converged do Algorithm 4 Dictionary Update
3: //Line 4-12 solving the problem of (L∗k+1 , S∗k+1 ) =
Input:
arg min L(L, S, Yk∗ , μk )
L,S Dic = [d1 ..., dk ] ∈ Rk×k
4: A0k+1 = A∗k , Ek+1
0
= Ek∗ , j = 0; B = [b1 ..., bk ] ∈ Rk×k
5: while not converged do A = [a1 ..., ak ] ∈ Rk×k
j
6: (U, S, V ) = svd(D − Ek+1 + u−1 ∗
k Yk );
Output:
j+1 T An updated Dic
7: Lk+1 = U Sμk−1 [S]V ;
1: Repeat
8: Sj+1 j+1 −1 ∗
k+1 = Sλμ−1 [D − Sk+1 + μk Yk ]; 2: for j = 1 to k do do
9: end while 3: Update the j-th column to optimize
10: ∗
Yk+1 = Yk∗ + μk (D − L∗k+1 − S∗k+1 ); μk+1 = ρμk ; 4: 1
uj ← A[j,j] (bj − Daj ) + dj
11: k = k + 1; 1
12: end while dj ← max(uj  ,l) uj
2
5: end for
6: Until convergence
and the sparse matrix of source data D is separated, the noise
of the source data are eliminated.
When the dictionary is trained, an Orthogonal Matching
C. Sparse Coding Pursuit (OMP) algorithm is used to get the sparse vector. In
Figure 4, all the data are used for the dictionary learning. Then
When the input source data are de-noised, a sparse coding when the dictionary is trained, the testing data and the training
algorithm with online dictionary learning is used. It extracts data are represented using the trained dictionary. The OMP
the core information of the de-noised data so that features of algorithm as it is the fastest resolution algorithm.
the data will be more separable. Each de-noised source data
li is a vector, that contains the information of the compressor.
D. The SVM classifier
As the sparse theory [10], each vector li , be decomposed into
two parts, the dictionary and the sparse vector: SVM classifier is used to classify the de-noised data. This
classifier uses a Gaussian radical basis function (RBF) kernel
li = Dicxi +λxi 1 , (10) [3] that is denoted in Equation 12.
where the Dic is a dictionary that contains common informa- 2
k(vi , vj ) = φ(vi )T φ(vj ) = exp(−γvi − vj  ) , (12)
tion of all the vectors li in L. And xi is a sparse feature vector
of li In that case, classify the corresponding feature vector can Where γ is the parameter of RBF widths. vi and vj represent
help us classify the input vector li . the ith and jth feature vector of the training images. φ is

76
) #  
 
    !
"  #
(  % $
   
 $

  #   #
   
 %  % 
Fig. 4. Sparse Representation of Reciprocating Compressor Data & $  & $ 

the infinite dimensional feature mapping function of the RBF & $'  & $' 
kernel.    
In this work, , an one-against-one strategy is applied to train
l(l − 1)/2 non-linear SVMs, where l is the number of faults * +  
categories. Given the training vectors vi ∈ Rn , i = 1, ..., s, in
tow classes, with class labels yi ∈ −1, 1. The SVM solves the  
following constraint convex optimization problem: 
s

2 Fig. 6. Data Analysis Workflow
min((1/2)ψ + C ξi )
i=1
, (13)
s.t.yi (ψ T φ(vi ) + b) ≥ 1 − ξi ξj > 0 suction valve reed broken, 2) suction valve leakage, 3) exhaust

s valve reed broken, 4) exhaust valve leakage, and 5) piston-rods
where C ξi ) is the regulation term for the non-linearly settling. The size of each group is 21 KB. The total size of
i=1
separable datasets. (vi ) + b) is the hyper-plane. The optimal ψ five million groups is about 100 TB.
should satisfy Equation 13. All experiments are done by using Matlab 2014a and Visual
Studio 2013 community edition mixed in a private cloud with
s
300 servers. Each server uses 8-core & 16-thread Intel Xron
 E5-2670 CPU with 16GB memory. The accuracy rate of fault
ψ= yi αi φ(vi ) , (14) diagnosis is the most important criterion and this is used for
i=1
evaluation.
where αi is the Lagrange multiplier. The decision function There are two steps in the experiment. The first step is data
for these two classes are expressed as follows pre-processing, and the second step is fault diagnosis. Figure
 s  6 shows the workflow of data analysis. In data pre-processing,

T the original data are divided into normal and faulty data. Then
sign(ψ φ(v) + b) = sign yi xi k(vi , v) + b , (15)
i=1
the faulty data are randomly picked as training and testing
data. The proposed solutions do fault analysis on training and
Using this one verses one SVM classifier, the de-noised testing data. The analyzed results of testing data are used to
data can be classified into a few groups. verify the analyzed results of training data. The verified results
are used in dictionary learning of training data.
IV. E XPERIMENTS AND A NALYSES
A. Data Pre-processing
Experiments are done to evaluate the proposed system
using data from China National Offshore Oil Corporation In data pre-processing, the original data are classified into
(CNOOC). The data are collected from operating compressors normal and faulty condition data. Only faulty condition data
in offshore oilfields, and contains five million groups of data are used to do fault diagnosis. According to the obtained
from different operation conditions. These conditions contain operating data, experiments are used to verify the validity and
normal operation conditions and five faulty conditions: 1) effectiveness of data pre-processing.

77
TABLE I. ACCURACY OF FAULT C ONDITION D ISCRIMINATION valve reed broken faults are successfully identified. Similarly,
Attempt 1 2 3 4 5 92.9% of exhaust valve leakage faults are diagnosed, and the
Accuracy(%) 99.3 100.0 98.3 97.5 96.9 rest 7.1% faults are diagnosed as suction valve leakage faults.
Attempt 6 7 8 9 10
Accuracy(%) 97.0 97.5 97.5 97.0 98.7
In this group experiment, piston-rods settling faults have the
lowest accuracy rate of fault diagnosis. Only 71.4% of piston-
rods settling faults are identified. 7.1%, 7.1% and 14.3% of
piston-rods settling faults are diagnosed as suction valve reed
broken faults, suction valve leakage faults and exhaust valve
reed broken faults respectively.
In the second group of experiments, all exhaust valve reed
broken faults are 100% identified. Only 61.5% of piston-
rods settling faults are diagnosed. The accuracy rates of fault
diagnosis in suction valve reed broken faults, suction valve
leakage faults, and exhaust valve leakage are 76.9%, 92.3%,
and 92.3% respectively. The fourth group experiment has the
best performance. Only 83.3% of suction valve leakage faults
 are identified. The rest 16.7% of suction valve leakage faults
are diagnosed as piston-rods settling faults. The other four
Fig. 7. Accuracy Rate of Fault Diagnosis types of faults are identified.
Figure 7 shows the results of ten-group experiments. The
In data pre-processing, half of normal condition and faulty accuracy rate of diagnosing each fault is marked in different
condition data are used as training data, and the rest of data color. In most cases, more than 85% faults can be identified.
are used as testing data to test the proposed system. Before Some types of faults have 100% accuracy rate of fault diagno-
doing fault diagnosis, the proposed system differentiates nor- sis. Only two types of faults have more than 60% accuracy rate
mal condition and faulty condition by using SVM classifier. of fault diagnosis in all experiments. The high accuracy rates
The same experiments are repeated ten times using different prove that the proposed fault diagnosis system can efficiently
training and testing data randomly selected from normal and diagnose faults from large data automatically. Figure 8 shows
faulty data. the analysis of fault-diagnosis results. The expected fault iden-
tification rate of five faults is 70%. Only fault identifications
The accuracy rates of ten attempts are shown in Table of piston-rods setting in 2nd attempt and exhaust valve reed
I. The second attempt has the highest accuracy rate 100%, broke in 6th attempt are lower than expectations. All other
and the fifth attempt has the lowest accuracy rate 96.9%. The results are better than expectations. In total 50 analyses (10
average accuracy rate is 97.97%, i.e. Almost 98% of faults can times for each fault), 25 analyses (50% of all analyses) have
be identified. The high accuracy of data identification increases 100% fault identification rate. The results of proposed system
the confidence in diagnosis. are better than expectation. The proposed system is efficient
to do fault diagnosis.
B. Fault Diagnosis
The proposed system is applied identify potential faults V. C ONCLUSION
after data pre-processing.
This paper proposed a fault-diagnosis system for compres-
Half of data (about 50 TB) are used for training, and the sors. The machine learning method is used to get a pattern to
rest (about 50 TB) for testing. In both procedures, RPCA does recognize working state and faulty state automatically. The
de-noise operation first. Then it does dictionary learning. After proposed system is evaluated using operational data from
that SVM classifier is applied to the learned data. Training oilfields, the results indicated that the system can identify most
results are used by testing process to get the final results. of these faults automatically with more than 80% accuracy.
Table II shows the dictionary learning time of training
data. In ten attempts, the 7th attempt has the maximum time R EFERENCES
139s and the 8th attempt has the minimum time 108s. For 50 [1] H. Abdi and L. J. Williams. Principal component analysis. Wiley In-
TB data size, the learning process is efficient. Ten groups of terdisciplinary Reviews: Computational Statistics, 2(4):433–459, 2010.
test results are shown in Figure 5 (1)-(10). In the first group [2] C. Chang and C. Lin. LIBSVM: A library for support vector machines.
experiment, 92.9% of suction valve reed broken faults are ACM TIST, 2(3):27, 2011.
successfully identified. The remaining 7.1% of suction valve [3] Y.-W. Chang, C.-J. Hsieh, K.-W. Chang, M. Ringgaard, and C.-J. Lin.
reed broken faults are identified as exhaust valve reed broken Training and testing low-degree polynomial data mappings via linear
faults by mistake. 100% of suction valve leakage and exhaust svm. J. Mach. Learn. Res., 11:1471–1490, Aug. 2010.
[4] M. Elhaj, F. Gu, A. Ball, A. Albarbar, M. Al-Qattan, and A. Naid.
Numerical simulation and experimental study of a two-stage recipro-
TABLE II. D ICTIONARY L EARNING T IME OF T RAINING DATA cating compressor forcondition monitoring. Mech. Syst. Signal Process,
22:374–389, 2008.
Attempt 1 2 3 4 5
Time 135s 122s 134s 121s 127s [5] K. M. Lee, Y. Matsushita, J. M. Rehg, and Z. Hu, editors. Computer
Attempt 6 7 8 9 10 Vision – ACCV 2012, volume 7724 of Lecture Notes in Computer
Time 137s 139s 108s 112s 119s Science. Springer, 2013.

78
[6] B. Lin, Q. Li, Q. Sun, M. Lai, I. Davidson, W. Fan, and J. Ye. Stochastic Symposium on Service Oriented System Engineering, SOSE 2014,
coordinate coding and its application for drosophila gene expression Oxford, United Kingdom, April 7-11, 2014, pages 59–69, 2014.
pattern annotation. CoRR, abs/1407.8147, 2014. [27] F. Zhang and Z. Shi. Optimal and adaptive battery discharge strategies
[7] Z. Lin, M. Chen, and Y. Ma. The augmented lagrange multiplier method for cyber-physical systems. In Decision and Control, 2009 held jointly
for exact recovery of corrupted low-rank matrices. arXiv preprint with the 2009 28th Chinese Control Conference. CDC/CCC 2009.
arXiv:1009.5055, 2010. Proceedings of the 48th IEEE Conference on, pages 6232–6237, Dec
[8] J. Mairal, F. R. Bach, J. Ponce, and G. Sapiro. Online learning for 2009.
matrix factorization and sparse coding. Journal of Machine Learning [28] F. Zhang, Z. Shi, and S. Mukhopadhyay. Robustness analysis for
Research, 11:19–60, 2010. battery-supported cyber-physical systems. ACM Trans. Embed. Comput.
[9] A. F. K. Morales and I. Mejı́a-Guevara. Evolutionary training of Syst., 12(3):69:1–69:27, Apr. 2013.
SVM for multiple category classification problems with self-adaptive
parameters. In Proceedings of 10th Ibero-American Conference on AI,
pages 329–338, 2006.
[10] I. Rish and G. Grabarnik. Sparse Modeling: Theory, Algorithms, and
Applications. CRC Press, Inc., Boca Raton, FL, USA, 1st edition, 2014.
[11] B. Shi and Y. Shen. The theory and practice of mechanical fault
diagnosis. Metallurgical Industry Press, 2000.
[12] B. N. S.S. Manepatil, G. Yadava. Modeling and computer simulation of
reciprocating compressor with faults. IE (I) Journal-MC, 81:108–116,
2000.
[13] J. Sun, Y. Chai, C. Su, Z. Zhu, and X. Luo. BLDC motor speed control
system fault diagnosis based on LRGF neural network and adaptive
lifting scheme. Appl. Soft Comput., 14:609–622, 2014.
[14] W. Tsai, C. J. Colbourn, J. Luo, G. Qi, Q. Li, and X. Bai. Test algebra
for combinatorial testing. In Proceedings of 8th IEEE International
Workshop on Automation of Software Test, AST 2013, San Francisco,
CA, USA, May 18-19, 2013, pages 19–25, 2013.
[15] W. Tsai, J. Luo, G. Qi, and W. Wu. Concurrent test algebra execution
with combinatorial testing. In Proceedings of 8th IEEE International
Symposium on Service Oriented System Engineering, SOSE, pages 35–
46, 2014.
[16] W. Tsai and G. Qi. DICB: dynamic intelligent customizable benign
pricing strategy for cloud computing. In Proceedings of 5th IEEE
International Conference on Cloud Computing, Honolulu, HI, USA,
June 24-29, 2012, pages 654–661, 2012.
[17] W. Tsai and G. Qi. Integrated adaptive reasoning testing framework
with automated fault detection. In Proceedings of IEEE International
Symposium on Service-Oriented System Engineering, SOSE 2015, San
Francisco Bay, CA, USA, March 30 - April 3, 2015, pages 169–178,
2015.
[18] W. Tsai, G. Qi, and Y. Chen. Choosing cost-effective configuration in
cloud storage. In Proceedings of 11th IEEE International Symposium on
Autonomous Decentralized Systems, ISADS 2013, Mexico City, Mexico,
6-8 March 2013, pages 1–8, 2013.
[19] W. Tsai, G. Qi, and K. Hu. Autonomous decentralized combinatorial
testing. In Proceedings of 12th IEEE International Symposium on
Autonomous Decentralized Systems, ISADS 2015, Taichung, Taiwan,
March 25-27, 2015, pages 40–47, 2015.
[20] W. Tsai, G. Qi, L. Yu, and J. Gao. Taas (testing-as-a-service) design
for combinatorial testing. In Proceedings of 8th IEEE International
Conference on Software Security and Reliability, SERE 2014, San
Francisco, California, USA, June 30 - July 2, 2014, pages 127–136,
2014.
[21] W.-T. Tsai, G. Qi, and Z. Zhu. Scalable saas indexing algorithms
with automated redundancy and recovery management. Int J Software
Informatics, 7(1):63–84, 2013.
[22] Y. Wang, C. Xue, X. Jia, and X. Peng. Fault diagnosis of reciprocating
compressor valve with the method integrating acoustic emission signal
and simulated valve motion. Mechanical Systems and Signal Process-
ing, 56-57:197–212, 2015.
[23] Wikipedia. Deepwater horizon oil spill. https://en.wikipedia.org/wiki/
Deepwater Horizon oil spill#Explosion/, 2016.
[24] Wikipedia. Piper alpha. https://en.wikipedia.org/wiki/Piper Alpha/,
2016.
[25] J. Wright, A. Ganesh, S. Rao, Y. Peng, and Y. Ma. Robust principal
component analysis: Exact recovery of corrupted low-rank matrices via
convex optimization. In Proceedings of 23rd Annual Conference on
Neural Information Processing Systems, pages 2080–2088, 2009.
[26] W. Wu, W. Tsai, C. Jin, G. Qi, and J. Luo. Test-algebra execution
in a cloud environment. In Proceedings of 8th IEEE International

79
˄˅ ˄˅

˄˅ ˄,˅

˄-˅ ˄.˅

˄/˅ ˄0˅

˄1˅ ˄2˅


Fig. 5. Fault Diagnosis Experiment Results

80

Fig. 8. Analysis of Fault-Diagnosis Results

81

You might also like