Sensor Fault Detection Based on Particle Filter and ...
Transcript of Sensor Fault Detection Based on Particle Filter and ...
Jordan Journal of Civil Engineering, Volume 13, No. 4, 2019
- 501 - © 2019 JUST. All Rights Reserved.
Sensor Fault Detection Based on Particle Filter and Mahalanobis Distance
Tianzhi Li 1), Gang Liu2) and Liangliang Zhang 3)
1) Master Student, School of Civil Engineering, Chongqing University, No.174 Shazhengjie, Shapingba, Chongqing, 400044, China. E-Mail: [email protected]
2) Professor, School of Civil Engineering, Chongqing University, No.174 Shazhengjie, Shapingba, Chongqing, 400044, China. E-Mail: [email protected]
3) Professor, School of Civil Engineering, Chongqing University, No.174 Shazhengjie, Shapingba, Chongqing, 400044, China. E-Mail: [email protected]
ABSTRACT
It is essential to evaluate the health condition of structure and sensors in structural health monitoring (SHM)
systems. Compared with the identification of structural damage, only a few papers investigated sensor fault
detection in civil engineering. This paper presents a model-based sensor fault detection approach utilizing
particle filter (PF) and Mahalanobis distance (MD). The discrete state-space model is constructed by a system
identification algorithm named N4SID instead of physical principles. A sensor fault is determined if
Mahalanobis distance between test state and training state is above the threshold. The experimental study
demonstrates the accuracy and efficiency of the proposed method.
KEYWORDS: Structural health monitoring, Sensor fault detection, Particle filter, Mahalanobis distance, System identification.
INTRODUCTION
During recent years, large-scale civil structures
equipped with multiple sensors are becoming popular as
a result of developing SHM systems. Since the structure
conditions are evaluated by the data measured by
sensors over the service period, it is essential to detect
the abnormal condition in sensors, which is called sensor
fault detection (FD). Sensor fault detection decides
whether there is an abnormal condition in the monitored
system (Gao, Cecati and Ding, 2015). In civil
engineering, a large list of relevant works are related to
the identification of structural damage. Only a few
papers focused on the detection of sensor faults (Kullaa,
2011; Li et al., 2018), which is solved in this study.
There are two major approaches for sensor fault
detection problems (Thomas, 2002): model-free and
model-based. In model-free methods, several physical
sensors at the same location measure the same quantity
to generate physical redundancy. The applications of
these methods are inevitably limited by many factors,
such as size, budget, weight,… etc. As to model-based
methods, a mathematical model of the monitoring
system is always employed to acquire redundancy.
Model-based methods with accurate models are able to
offer better performance compared with model-free
methods (Chen and Patton, 1999).
Generally, the residual between the measured output
from sensor and estimated output from the model is used
as detection feature for most model-based FD problems.
The residual generation can be broadly classified as
(Poon et al., 2016): observer -based methods, parameter
estimation methods or statistical approaches. Among Received on 3/12/2018. Accepted for Publication on 5/10/2019.
Sensor Fault Detection… Tianzhi Li, Gang Liu and Liangliang Zhang
- 502 -
these approaches, the observer-based methods,
including Kalman filter, particle filter, multiple model
filter,… etc, are the most investigated fields in sensor
FD problems (Khan et al., 2017). These approaches are
also most promising due to their flexibility in choosing
proper observers to match performance requirements
(Cai and Wu, 2011).
In this paper, therefore, the observer-based approach
is selected and we apply the powerful algorithm in a
nonlinear system called particle filter. The residuals of
estimated output and measured output based on particle
filter (PF) are extensively employed in numerous studies
(V. Kadirkamanathan et al., 2002; Duan et al., 2008;
Tadić and Ðurović, 2014). Meanwhile, based on the fact
that the sum of unnormalized particle weights declines
if a sensor fault occurs, the decision rule can also be
developed by the weight of each particle (Li and
Kadirkamanathan, 2001; Yin and Zhu, 2015). However,
in a highly nonlinear system, the one-dimensional
residual or weight is usually unable to offer
comprehensive evaluation for system health condition.
For more accurate sensor FD results, PF is then
exploited to generate the state vector to construct the
relatively precise detection function (Wei, Huang and
Chen, 2009). Inspired from above, state vector is
selected as sensor FD index and then Mahalanobis
distance, powerful in measuring the correlation of data,
is applied to develop sensor FD rule.
The other parts of the paper are organized as follows:
firstly, the sensor FD problem is formulated; then, the
background of particle filter and Mahalanobis distance
is introduced; following that, the detection and
identification approach is described in detail; finally, the
experimental example illustrates the effectiveness of
this proposed method in sensor FD problems.
PROBLEM FORMULATION
In model-based methods, the model is obtained by
using either physical principles or system identification
techniques. It is difficult to construct a state-space model
of a large-scale civil structure by physical principles and
measure precise excitation, so the output-only space-
state model is established using system identification in
this study.
𝒙 𝑨𝒙 𝒘 (1)
𝑦 𝑯𝒙 (2)
where x and 𝑦 represent the state vector and true
output vector; A is the system state transmission with
nn dimension; H is the output influence matrix with
1n dimension; n is the dimension of the model; w is
the system noise vector. The subscripts k and k+1
describe the step time.
In the absence of sensor faults, the measured output
vector yk is described as 𝑦 𝑦 𝑣 , where v is the
measurement noise vector. Otherwise, there are
residuals between yk and 𝑦 . Three common types of
sensor faults are considered here, such that gain 𝑦𝑎 𝑦 𝑣 , bias 𝑦 𝑦 𝑏 𝑣 and constant 𝑦𝑐 𝑣 , where a, b, c are invariant constants, the
superscript i denotes the fault sensor channel. More
details about those three types of sensor faults can be
found in the reference (Li et al., 2018).
FAULT DETECTION AND ISOLATION
Background on Particle Filter
As a powerful method for the construction of
approximate probability density functions
(PDF) 𝑝 𝒙 |𝒚 : in linear and nonlinear systems, the
procedure of particle filter algorithm is expressed as
follows (Gordon, Salmond and Smith, 2002):
Predict: there are a set of particles 𝒙 : 𝑖 1, … , 𝑁
from known PDF 𝑝 𝒙 |𝒚 : and then a set of
samples 𝒙 ∗: 𝑖 1, … , 𝑁 is obtained by Equation
(1), where N is the amount of particles.
Update: considering measurements yk, a weight 𝑞 is
assigned to each 𝒙 ∗:
𝑞|𝒙 ∗
∑ |𝒙 ∗ (3)
Jordan Journal of Civil Engineering, Volume 13, No. 4, 2019
- 503 -
A discrete distribution over 𝒙 ∗: 𝑖 1, … , 𝑁 is
defined and the weight 𝑞 is related to state vector 𝒙 ∗.
Resample: a set of values 𝒙 : 𝑖 1, … , 𝑁
obtained by the above distribution 𝒙 ∗: 𝑖 1, … , 𝑁
after resampling finally construct approximate PDF
𝑝 𝒙 |𝒚 : .
𝑝 𝒙 |𝒚 : ∑ 𝑞 𝛿 𝒙 𝒙 (4)
Once p(xk|y1:k) is determined, we can develop the
minimum mean squared error (MMSE) estimation of the
state xk.
𝒙 𝒙 𝑝 𝒙 |𝒚 : 𝑑𝒙 (5)
If a fault arises, the change of measured output leads
to significant discrepancy between the state vector 𝒙
and 𝒙 .
Sensor Fault Detection Approach
There are four types of data utilized in this study;
namely, N4SID data y(n), training data y(r), threshold data
y(h) and test data y(t). The first three data types are
obtained in baseline condition, when the sensor is
healthy. We apply N4SID data to construct the state-
space model. It is worth noting that there will be an
individual output-only state-space model for each
sensor. Then, training data is applied to generate the
state vector 𝒙 by Equation (5). A fix-width (m=300)
sliding window of state ∑ 𝒙 serves as fault index:
∑ 𝒙 𝒙 𝒙 ⋯ 𝒙 (6)
In order to construct the threshold, the state matrix X
of training data formulated in Equation (7) is applied to
generate its mean vector μ and matrix covariance S.
𝑿 ∑ 𝒙 ∑ 𝒙 … ∑ 𝒙 (7)
The sum of state vector ∑ 𝒙 or ∑ 𝒙
formed by the threshold data or test data, μ and S,
constructs Mahalanobis distance.
𝑀𝐷 ∑ 𝒙 𝝁 𝑺 ∑ 𝒙 𝝁 (8)
𝑀𝐷 ∑ 𝒙 𝝁 𝑺 ∑ 𝒙 𝝁 (9)
Then, 𝑀𝐷 : 𝑗 1,2 … 𝑁 are sorted in a
descending order and the control limit h chosen is
determined.
ℎ 𝑀𝐷 ∗ (10)
The empirical coefficient α is selected as 0.01
according to the reference (Verdier and Ferreira, 2011).
A fault is detected while 𝑀𝐷 formed by Equation
(8) is above the threshold h. All the sensor fault
detection procedure is presented in Figure 1.
Figure (1): Sensor fault detection procedure
Training DataOutput-onlyState-space
Model
N4SID Data
N4SID
Threshold Data
μ S
Test Data
0ˆm h
k iix
0ˆm t
k iix
Threshold h
There is a fault
Yes
tkMD
?tkMD h
Sensor Fault Detection… Tianzhi Li, Gang Liu and Liangliang Zhang
- 504 -
EXPERIMENTAL STUDY
Experimental Details
In order to assess the performance of the proposed
approach in reality, a two-floor frame described in
Figure 2 was built in the laboratory. The acceleration of
each floor was measured at X direction with a sampling
frequency of 200Hz. There were four circle columns
with 7mm diameter on each floor. Data representing
sensor faults was artificially created by healthy data. We
assumed three typical sensor faults; namely, gain, bias
and constant, to have occurred in sensor 1 and sensor 2.
The acceleration data of the two sensors was
collected under environmental noise, which meant that
the structural dynamic behavior was mainly controlled
by stochastic excitation and the signals should remain
relatively steady all the time. The 90s data (18000 steps)
presented in Figure 3 was under baseline condition when
the two sensors were healthy. We defined the data from
time step 1 to 3000 (15s) as N4SID data, data from time
step 3001 to 6000 (15s) as training data, data from time
step 10001 to 13000 (15s) as threshold data and data
from time step 15001 to 18000 (15s) as test data. We
assumed that there was one faulty sensor at one time and
data representing sensor faults was artificially created by
the healthy test data (Case 1). The seven test cases listed
in Table 1 were included to verify the feasibility of this
FD rule.
Figure (2): Experimental equipment
Figure (3): Healthy data utilized in this study
Jordan Journal of Civil Engineering, Volume 13, No. 4, 2019
- 505 -
Table 1. Seven test cases
Test Case Faulty Sensor Faulty Type Parameter 1 All healthy2
1 Gain a=1.5
3 Bias b=0.0003 4 Constant c=0.0003 5
2 Gain a=1.5
6 Bias b=0.0003 7 Constant c=0.0003
The N4SID algorithm, widely used for system
identification and easily achieved by MATLAB system
identification box (Ljung, 1991), formed the optimal
parameters of A, H, w and v by the N4SID data. It is
worth noting that there is an individual state-space
model for each sensor. The training data from step 3001
to 6000 was employed to generate μ, S. Then, the
threshold data between time step 10001 and 13000
provided ∑ 𝒙𝒌𝒉 to construct the threshold by
Equation (8). Figure 4 presents 2700 samples
𝑀𝐷 : 𝑘 1,2, … ,2700 and the control limits
h1=8.18, h2=6.20 calculated by Equation (10). The
thresholds are applied to FD.
Figure (4): Mahalanobis distances and control limits calculated by threshold data
Sensor Fault Detection Results ∑ 𝒙 : 𝑘 15001,17002, … ,17700 from test
data and μ, S were applied to calculate Mahalanobis
distances 𝑀𝐷 : 𝑘 1,2, … ,2700 by Equation (9).
Figure 5 and Figure 6 represent the detection
performance for the three types of faults in sensor 1 and
sensor 2, respectively. In case 1, all Mahalanobis
distances are below the threshold, which means that
there is no sensor fault. As to the other test case, sensor
faults, including gain, bias and constant, are successfully
detected, since their MDs are above the threshold.
Moreover, this fault detection method shows its
sensitive detection ability to bias and constant, since the
MDs in cases 3, 4, 6 and 7 are largely above the
thresholds.
Sensor Fault Detection… Tianzhi Li, Gang Liu and Liangliang Zhang
- 506 -
Figure (5): Fault detection in sensor 1
Figure (6): Fault detection in sensor 2
CONCLUSION
A sensor fault detection approach for SHM systems
is presented in this paper. PF is employed to generate
system states of training data and test data. Then,
Mahalanobis distance is utilized to detect sensor faults.
Three typical sensor faults; namely, gain, bias and
constant, are used to verify the proposed method. The
results show that this method is very sensitive to these
faults, especially bias and constant. The discrimination
of structural damage and sensor faults is not investigated
in this paper. Our future work intends to solve this
problem.
ACKNOWLEDGEMENT
The authors would like to thank the support from the
Natural Sources Foundation of China (Grant No.
51578095).
Jordan Journal of Civil Engineering, Volume 13, No. 4, 2019
- 507 -
REFERENCES
Cai, Xuejing, and Fen Wu. (2011). "Multiobjective fault
detection and isolation for flexible air-breathing
hypersonic vehicle". Journal of Systems Engineering
and Electronics, 22, 52-62.
Chen, Jie, and Ron J. Patton. (1999). "Robust model-based
fault diagnosis for dynamic systems". (Springer, US).
Duan, Zhuo Hua, ZiXing, Yu, and Jin Xia. (2008). "An
adaptive particle filter for soft fault compensation of
mobile robots". Science in China, 51, 2033-2046.
Gao, Zhiwei, Carlo Cecati, and Steven X. Ding. (2015). "A
survey of fault diagnosis and fault-tolerant techniques-
part i: fault diagnosis with model-based and signal-based
approaches". IEEE Transactions on Industrial
Electronics, 62, 3757-3767.
Gordon, N. J., D. J. Salmond, and A. F. M. Smith. (2002).
“Novel approach to nonlinear/non-Gaussian Bayesian
state estimation”. IEE Proceedings F - Radar and Signal
Processing, 140, 107-113.
Khan, Rudaba, Paul Williams, Paul Riseborough, Asha Rao,
and Robin Hill. (2017). “Fault detection and
identification - a filter investigation”. International
Journal of Robust and Nonlinear Control, 28.
Kullaa, Jyrki. (2011). “Distinguishing between sensor fault,
structural damage and environmental or operational
effects in structural health monitoring”. Mechanical
Systems and Signal Processing, 25, 2976-2989.
Li, Lili, Gang Liu, Liangliang Zhang, and Qing Li. (2018).
“Deep learning-based sensor fault detection using s-long
short-term memory networks".
Li, Ping, and Visakan Kadirkamanathan. (2001). “Particle
filtering based likelihood ratio approach to fault
diagnosis in nonlinear stochastic systems”. IEEE
Transactions on Systems, Man and Cybernetics, Part C
(Applications and Reviews), 31, 337-343.
Ljung, Lennart. (1991). "System identification toolbox for
use with MATLAB: user's guide (Math Works)".
Poon, Jason, Palak Jain, Ioannis C. Konstantakopoulos,
Costas Spanos, Sanjib Kumar Panda, and Seth R.
Sanders. (2016). “Model-based fault detection and
identification for switching power converters”. IEEE
Transactions on Power Electronics, 32, 1419-1430.
Tadić, Predrag, and Željko Ðurović. (2014). “Particle
filtering for sensor fault diagnosis and identification in
nonlinear plants”. Journal of Process Control, 24, 401-
409.
Thomas, Philippe. (2002). “Fault detection and diagnosis in
engineering systems". Janos J. Gertler; Marcel Dekker,
Inc., New York, 1998, ISBN 0-8247-9427-3”. Control
Engineering Practice, 10, 1037-1038.
V. Kadirkamanathan, P.Li, M. H. Jaward, and S. G. Fabri.
(2002). “Particle filtering-based fault detection in non-
linear stochastic systems”. International Journal of
Systems Science, 33, 259-265.
Verdier, Ghislain, and Ariane Ferreira. (2011). “Adaptive
Mahalanobis distance and $ k $-nearest neighbor rule for
fault detection in semiconductor manufacturing”. IEEE
Transactions on Semiconductor Manufacturing, 24, 59-
68.
Wei, Tao, Yufei Huang, and C. L. Philip Chen. (2009).
"Adaptive sensor fault detection and identification using
particle filter algorithms". (IEEE Press).
Yin, Shen, and Xiangping Zhu. (2015). “Intelligent particle
filter and its application to fault detection of nonlinear
systems”. IEEE Transactions on Industrial Electronics,
62, 3852-3861.