Professional Documents
Culture Documents
Cjme 2014 03 475
Cjme 2014 03 475
Sensor Configuration and Test for Fault Diagnoses of Subway Braking System
Based on Signed Digraph Method
Received January 16, 2013; revised January 7, 2014; accepted April 1, 2014
Abstract: Fault diagnosis of various systems on rolling stock has drawn the attention of many researchers. However, obtaining an
optimized sensor set of these systems, which is a prerequisite for fault diagnosis, remains a major challenge. Available literature
suggests that the configuration of sensors in these systems is presently dependent on the knowledge and engineering experiences of
designers, which may lead to insufficient or redundant development of various sensors. In this paper, the optimization of sensor sets is
addressed by using the signed digraph (SDG) method. The method is modified for use in braking systems by the introduction of an
effect-function method to replace the traditional quantitative methods. Two criteria are adopted to evaluate the capability of the sensor
sets, namely, observability and resolution. The sensors configuration method of braking system is proposed. It consists of generating
bipartite graphs from SDG models and then solving the set cover problem using a greedy algorithm. To demonstrate the improvement,
the sensor configuration of the HP2008 braking system is investigated and fault diagnosis on a test bench is performed. The test results
show that SDG algorithm can improve single-fault resolution from 6 faults to 10 faults, and with additional four brake cylinder pressure
(BCP) sensors it can cover up to 67 double faults which were not considered by traditional fault diagnosis system. SDG methods are
suitable for reducing redundant sensors and that the sensor sets thereby obtained are capable of detecting typical faults, such as the
failure of a release valve. This study investigates the formal extension of the SDG method to the sensor configuration of braking system,
as well as the adaptation supported by the effect-function method.
Keywords: fault diagnosis, subway braking system, signed directed graph, sensor set optimization.
them, developed by UMEDA, et al[7], is based on SDG modeling. As indicated in Fig. 1, the arc from S to C
quantitative models of tested systems. The other one, is noncausal because a deviation in S will cause a
proposed by LAPP, et al[8], uses flow sheets to develop corresponding fluctuation in C, whereas a change in C
SDG models. The former requires specific formulas of would not necessarily affect S.
different variables, whereas the latter may produce complex
SDG models that are difficult to handle. Therefore, a
method based on the effect function is introduced to
facilitate SDG modeling[9].
This paper is organized as follows. Following this
introduction, section 2 presents a brief description of the
sensor configuration based on SDG modeling using effect
functions. Section 3 introduces the application of the SDG
method to a case study of the HP2008 braking system.
Subsequently, in section 4, we present and analyze the
results of detailed bench tests conducted on the optimized
sensor set obtained by the SDG method. Finally, we offer
some concluding remarks in section 5.
2 Effect-function Modeling and SDG-based Fig. 1. Typical SDG model including a control loop
Sensor Configuration
In the chemical industry, most process systems can be
The design of an optimized senor configuration can be modeled as a system of differential equations (DE),
broadly divided into two tasks: fault modeling and sensor algebraic equations (AE), or differential algebraic
location identification. For most control systems, fault equations (DAE). The algorithms for DE, AE, and DAE
modeling requires specific knowledge of fault propagation systems used to facilitate the modeling process have
or cause-effect behavior, in which SDG-based causal already been studied by MAURYA, et al[10]. However,
models are widely utilized because of their advantages of mathematical equations are sometimes difficult to derive,
not only capturing the information flow but also its especially for designing fault diagnosis systems for railway
direction [10]. After the completion of fault modeling, the brakes. Thus, a method originally developed by LI, et al[9]
optimized sensor configuration can be obtained by to simplify the SDG modeling process is considered. Based
generating bipartite graphs based on different design on the modeling requirement of railway braking systems,
criteria, such as the observability, resolution, and the efficiency of the original modeling method is improved.
reliability. The detailed modeling procedure is as follows:
(1) Divide a control system into several units.
2.1 SDG modeling using effect functions (2) Define the variables of each unit.
When diagnosing complex systems, there is the problem (3) Obtain the effect functions concerning the
of insufficiently accurate mathematical models that can be relationship of the variables. In effect functions, all the
used to adequately represent the systems. Borrowing the variables are listed separately on the left, and the variables
concept of graph theory, the SDG can extract the essence of affecting them are listed on the right with signs concerning
the causal relationships in a given system. Generally, a the direction of the effect.
typical SDG model includes nodes (corresponding to (4) Obtain the SDG model of each unit.
variables) and arcs (representing the causal influences (5) Combine unit models into the system model of the
between the nodes). The influences are represented by the control system.
positive and negative arcs because the cause and effect (6) Obtain the system fault model by adding fault nodes
variables tend to change in the same or opposite direction. to the system model.
Because SDG models are purely qualitative, the legal
values that can be assigned to the nodes are restricted to the 2.2 SDG-based sensor configuration
set {, 0, }, where - indicates low deviation, 0 indicates Once a system fault model is obtained, the sensor
normal deviation, and indicates high deviation[11]. configuration based on it can be analyzed by following two
Furthermore, it is assumed that fault nodes are only allowed steps.
to be in the “0” or “” state, meaning that the (1) Generate different bipartite graphs of the SDG
representation of the signs are handled by the signed arcs system fault model according to different optimization
originating from the fault nodes. Because most systems are criteria.
closed-loop control systems, which tend to eliminate (2) Solve the set cover problem given by the bipartite
disturbances occurring at controlled variables, noncausal graphs and obtain the optimized sensor sets for each
arcs are defined to characterize the control loops during optimization criterion.
CHINESE JOURNAL OF MECHANICAL ENGINEERING ·477·
Bipartite graph generation is the kernel process of the can see from the figure, the approach focuses on searching
SDG algorithm because it transfers problems with different the unmarked sensors with highest unmarked arcs, which is
optimization criteria to a singular form that can be further a typical characteristic of a greedy algorithm[15]. Because
processed by a greedy algorithm. The generation methods greedy algorithms may keep some redundant sensors, the
differ according to their optimization criteria and the marked sensors without any arcs are deleted.
number of faults considered. In this paper, four generation
methods with four different criteria are considered:
observability with single-fault assumption, observability
with double-fault assumption, fault resolution with
single-fault assumption, and fault resolution with
double-fault assumption. Observability requires every
defined fault to be observed by at least one sensor, whereas
resolution refers to the ability to identify the exact fault that
has occurred. Other criteria include reliability, robustness,
and so forth[12–13]. A simplified schematic of bipartite graph
generation for single-fault observability and resolution is
shown in Fig. 2. When observability is considered, a SDG
model can be replaced by a digraph (DG) model because
the signs are no longer important. In the resulting DG
model, one needs to find the measurable nodes affected by
each fault, and then, the bipartite graph is formed. When
resolution is considered, the signs cannot be neglected
anymore, and set theory is used to construct the node sets
that contain the information crucial to distinguish between
faults. For instance, Bdc+ and Bdc- are nodes created to Fig. 3. Greedy algorithm for finding minimal sensor sets
make a distinction between D and C. The extended details
of the four methods are rather lengthy and thus will not be
covered in this paper. 3 Application to HP2008 Braking System
After the system units are modeled, the SDG model for
the entire braking system is created. This is done using the
effect functions representing the relation between different
units:
BCP ROP,
(3)
RPP BCP.
The possible faults of each component in HP2008 Table 2. Affected nodes under the single-fault observability
braking system in a service brake process are listed in assumption
Table 1. Because the channels within the BCU are created Affected sensor Affected sensor
Fault node Fault node
by drilling inside an air-circuit board, the occurrence rate of nodes nodes
leakage is quite low[16]. Therefore, it is proper to focus on CPi [EPC] BCP4i [AC4]
ROPi [φ] RS1i [AC1, BCP1]
the failures of the valves and sensors. To facilitate the
BCP1i [AC1] RS2i [AC2, BCP2]
application of the SDG algorithm, all the component BCP2i [AC2] RS3i [AC3, BCP3]
failures obtained from the users’ feedback are represented as BCP3i [AC3] RS4i [AC4, BCP4]
fault nodes, which can be conveniently added to the SDG
model. Thus, the SDG model for the calculation of sensor sets Under the double-fault assumption, the affected sensor
using these fault nodes can be created, as shown in Fig. 7. nodes for every fault node are listed in Table 3. Using these
data, the calculated sensor set for double-fault observability
Table 1. Fault nodes and variable nodes is {EPC, AC1, AC2, AC3, AC4, BCP1, BCP2, BCP3,
Variable BCP4}, whereas the actual sensor set considering the
Fault node Possible reason
node control sensors is {EPC, CS, AC1, AC2, AC3, AC4, BCP1,
CPi Brake valve leakage BCP2, BCP3, BCP4, RSS1, RSS2, RSS3, RSS4}.
CP
CPi Release valve leakage
CS CSF Dysfunctional CV pressure sensors Table 3. Affected nodes under double-fault observability
EPV EPVF Dysfunctional EP valves assumption
RV RVF Dysfunctional relay valve
Affected sensor Fault Affected
ROPi Relay valve leakage Fault nodes
ROP nodes nodes sensor nodes
ROPi Relay valve leakage
[ROP, AC1, (BCP2i,
BCP BCPi Brake cylinder leakage (CPi, CSF) [BCP2]
AC2, AC3, AC4] RSS2F)
RS RSi Skidding wheelsets
[ROP, AC1, (BCP3i,
RSS RSSF Dysfunctional rotational speed sensor (ROPi, RVF) [BCP3]
AC2, AC3, AC4] RSS3F)
ASV ASVF Dysfunctional antiskid valve
[AC1, BCP1, ROP, (BCP4i,
(RS1i, RVF) [BCP4]
AC2, AC3, AC4] RSS4F)
[AC2, BCP2, ROP, (RS1i,
(RS2i, RVF) [φ]
AC1, AC3, AC4] RSS1F)
[AC3, BCP3, ROP, (RS2i,
(RS3i, RVF) [φ]
AC1, AC2, AC4] RSS2F)
[AC4, BCP4, ROP, (RS3i,
(RS4i, RVF) [φ]
AC1, AC2, AC3] RSS3F)
(BCP1i, (RS4i,
[BCP1] [φ]
RSS1F) RSS4F)
Under the double-fault assumption, the SDG algorithm is Interconnect) data board, three BECUs, three BCUs, and a
much more complicated, and thus, only a part of the vibration platform. The vibration platform is utilized to
affected nodes is listed in Table 5. Using the greedy mimic the actual condition inside subway trains. The
algorithm, the sensor set for double-fault resolution will be fundamental advantages of using the test bench capable of
{EPC, BCP1, BCP2, BCP3, BCP4, RSS1, RSS2, RSS3, simulating most faults in subway braking systems are as
RSS4, AC1, AC2, AC3, AC4} with the actual set being follows.
{EPC, CS, BCP1, BCP2, BCP3, BCP4, RSS1, RSS2, RSS3, (1) Using the test bench, the tests can be carried out
RSS4, AC1, AC2, AC3, AC4}. repetitively and with most environmental variables
controlled.
Table 5. Affected nodes under double-fault resolution (2) It is more convenient and economical to add or
assumption reduce sensors in a test bench than in a real braking system
Fault nodes Affected sensor nodes used by Shanghai Metro.
(CPi, EPVF) [CS, EPC, ROP, AC1, AC2, AC3, AC4]
[CS, EPC, ROP, BCP1, BCP2, BCP3,
(CPi, EPVF) BCP4, RSS1, RSS2,RSS3, RSSR,
AC1, AC2, AC3, AC4]
(CPi, CSF) [ROP, AC1, AC2, AC3, AC4]
[ROP, BCP1, BCP2,BCP3, BCP4, RSS1,
(CPi, CSF) RSS2, RSS3, RSSR, AC1, AC2, AC3,
AC4]
(ROPi, RVF) [ROP, AC1, AC2, AC3, AC4]
[ROP, BCP1, BCP2, BCP3, BCP4, RSS1,
(ROPi, RVF) RSS2, RSS3, RSSR, AC1, AC2, AC3,
AC4]
[AC1, BCP1, ROP, BCP2, RSS2, AC2,
(RS1i, RVF) Fig. 8. Air brake test bench
BCP3, RSS3, AC3, BCP4, RSS4, AC4]
4.2 Test results equivalent sensor numbers for both situations are equal to
As previously mentioned, the SDG algorithm for sensor 10 and 14, respectively. The difference between the general
optimization is qualitative, and after the optimization sensors and tangible ones is that the former type treats the
process, the engineers must design detailed fault diagnosis values inside the controllers as sensor values.
algorithms or rules. Therefore, it is of the utmost Furthermore, there is a huge gap between the counted
importance to verify whether the optimized sensor set can double-fault and detectable fault because current fault
truly observe the faults used during the optimization detection methods in subway braking systems are generally
process. We verified this by simulating these faults using rule based[18]. Because most rules currently utilized are
the air brake test bench. As the details of all these faults can purely empirical, their number is quite limited. If the gap
be overwhelming, a specific example concerning release can be filled through the introduction of new algorithms
valve failure is described elaborately. Fig. 10 shows that a capable of detecting faults systematically, the accuracy of
normal staged brake is always characterized by alternating subway brake fault diagnosis will be greatly improved. As
inflation and deflation. Nevertheless, the deflation illustrated in Fig. 11(b), the original diagnosis method can
phenomenon disappears when the release valve is broken only detect six single faults of the HP2008 braking system.
and can no longer connect to the atmosphere. According to After the application of the SDG method, we increase the
the affected nodes in Table 5, the sensors affected by the number of detectable single faults to 10 faults and include
release valve failure include CS, ROP, and BCPi, 67 additional double faults that were not detectable before.
which is verified by Fig. 10, as all these pressures are
comparatively higher than their counterparts during a
normal staged brake. By studying all the fault patterns, it is
proven that the affected sensor nodes used in the SDG
algorithm qualitatively predict the deviation of the state
variables during system failures.
graphs based on criteria such as observability or resolution. [9] LI A, XIA T, ZHANG B, et al. SDG modeling approach for
To prove the efficiency of our approach, a case study using chemical engineering process[J]. Journal of System Simulation,
2003, 15(10): 1364–1368.
the HP2008 braking system is introduced in detail, and the
[10] MAURYA M R, RENGASWAMY R, VENKATASUBRAMANIAN
bench tests using the air brake test bench enabled us to V. Application of signed digraphs-based analysis for fault diagnosis
make the following conclusions. of chemical process flowsheets[J]. Engineering Applications of
(1) The SDG approach is capable of generating Artificial Intelligence, 2004, 17(5): 501–518.
optimized sensor sets. Regarding the HP2008 braking [11] OUASSIR M, MELIN C. Causal graphs and rule generation:
application to fault diagnosis of dynamic processes[C]//Proceedings
system, the sensor set solution is proposed.
of the 10th International Conference on Industrial and Engineering
(2) For the HP2008 braking system, we find that the Applications of Artificial Intelligence and Expert Systems, Goose
SDG algorithm can improve the single-fault resolution Pond Press, October 28, 1997: 367–373.
from 6 faults to 10 faults, and with an additional four BCP [12] ALI Y, NARASIMHAN S. Sensor network design for maximizing
sensors it can cover up to 67 double faults, which were not reliability of linear processes[J]. AIChE Journal, 1993, 39(5):
considered by the traditional fault diagnosis system. 820–828.
[13] BHUSHAN M, NARASIMHAN S, RENGAWAMY R. Robust
(3) The potential faults of the air braking system can be
sensor network design for fault diagnosis[J]. Computers and
detected by the sensor sets calculated using this method. It Chemical Engineering, 2008, 32(4–5): 1067–1084.
is useful to build the fault map and diagnose them in detail. [14] LOZIN V V. On maximum induced matchings in bipartite graphs[J].
Information Processing Letters, 2002, 81(1): 7–11.
References [15] KODAGANALLUR V, SEN A K. Greedy by chance – stochastic
[1] BARBOZA D. Shanghai subway accident injures hundreds[N/OL]. greedy algorithms[C]//Sixth International Conference on Autonomic
The New York Times, (2011) [2011-9-28]. http://www.nytimes.com/ and Autonomous Systems (ICAS), Cancun, Mexico, IEEE, March
2011/09/28/world/asia/shanghai-subway-accident-injures-hundreds. 7–13, 2010: 182–187.
html. [16] YU Z, CHEN D. Modeling and system identification of the braking
[2] FRANK P M. New developments using AI in fault diagnosis: system of urban rail vehicles[J]. Journal of the China Railway
engineering applications[J]. Artificial Intelligence, 1997, 10(1): Society, 2011, 33(10): 37–40.
3–14. [17] WANG X, WU M, CHEN Z. Application with theory of vibration
[3] VAZQUEZ J, MAZO M, LAZARO J L, et al. Detection of moving equivalence on reliability test bed of train brake system[J]. Applied
objects in railway using vision[C]//IEEE Intelligent Vehicles Mechanics and Materials, 2010, 34(35): 1978–1982.
Symposium, Parma, Italy, June 14–17, 2004: 872–875. [18] DALEY S, NEWTON D A, BENNETT S M, et al. Methods for
[4] ROBERTS C, GOODMAN C, BAI H. A generic fault detection and fault diagnosis in rail vehicle traction and braking systems[C]//IEE
diagnosis approach for railway assets[C]//International Conference Colloquium on Qualitative and Quantitative Modelling Methods for
on Railway Engineering – Challenges for Railway Transportation Fault Diagnosis, London, England, April 24,1995: 501–513.
in Information Age, Hong Kong, China, March 25–28, 2008: 1–8.
[5] LI P, KONG F, DANG L. A new fault diagnosis system for train Biographical notes
bearings based on PCA and ACO[C]//Logistic Systems and ZUO Jianyong, born in 1976, is currently an associate professor at
Intelligent Management, Harbin, China, January 9–10, 2010: 526–530. Institute of Railway and Urban Mass Transit, Tongji University,
[6] BHUSHAN M, RENGASWAMY R. Design of sensor network China. He received his PhD degree from Shanghai Jiaotong
based on the signed directed graph of the process for efficient fault University, China, in 2005. His research interests include
diagnosis[J]. Industrial and Engineering Chemistry Research, 2000, simulation and control of trains’ braking systems.
39(4): 999–1019. Tel: +86-21-69584712; E-mail: zuojy@tongji.edu.cn
[7] UMEDA T, KURIYAMA T, O’SHIMA E, et al. A graphical
approach to cause and effect analysis of chemical processing CHEN Zhongkai, born in 1987, is currently an M.S. candidate at
systems[J]. Computer Applications in Chemical Engineering. Institute of Railway and Urban Mass Transit, Tongji University,
Proceedings of the 12th European Symposium, 1980, 35(12): China. He received his bachelor’s degree from Tongji University,
2379–2388. China, in 2010. His main research interests include artificial
[8] LAPP S A, POWERS G J. Computer-aided synthesis of intelligence and control theory.
fault-trees[J]. IEEE Transactions on Reliability, 1977, 26(1): 2–13. Tel: +86-21-69584712; E-mail: craigknox@126.com