You are on page 1of 8

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Data-driven Robust Fault Detection and Isolation


of Three Phase Induction Motor

Abstract—This paper deals with data-driven fault diag- model due to neglected dynamics. So, in this work, data-
nosis of three phase induction motor. The tools from sub- driven fault diagnosis of induction motor is presented. Here,
space identification based data driven scheme are utilized no system model required for design of FD scheme for motor.
to design a framework of fault detection and isolation.
The fault detection scheme is so designed that it caters Some multivariate data-driven FD techniques have also been
for even small magnitude faults. In order to discriminate applied to induction motor. In [4], Principle Component
different faults, a fault isolation algorithm is proposed. Analysis was used for dimensionality reduction of data and
Isolation thresholds are designed. The results show the then Kernel Density Estimation is used to obtain patterns
effectiveness of the proposed schemes. representing the faults. In [5], bearing fault diagnosis was
Index Terms—Data-driven, fault detection, fault isolation, performed such that optimal features were selected from
fault diagnosis, three phase induction motor data of induction motor. These multivariate data-driven
schemes are based on complex statistical approaches unlike
Subspace Identification Method (SIM), which is based on
I. I NTRODUCTION
concepts of linear algebra. So, for designing FD system

T HREE phase induction motor plays a vital role in an in-


dustrial process. It is desired that it operates without any
fault. However, if it gets faulty, then fault should be detected
for induction motor, we apply the SIM based data-driven
FD scheme. Its benefits include simple implementation and
usage of model based FD techniques such as parity based
as quickly as possible to avoid any serious damage to the residual or diagnostic observer unlike above described
induction motor or other associated equipment. Various types MVA techniques. Moreover, due to advanced SIM algo-
of faults can occur in induction motor that are 1) Stator faults rithm in which parity vector can be directly identified from
which include insulation rupture of stator winding and inter- input-output data, without identifying the system model
turn short circuit faults which may result into damaging of [6], relatively lesser computational effort is required by
stator core or burning of stator winding 2) Sensor faults caused SIM based FD scheme than MVA schemes.
by abnormalities of measurement devices that include bias Our work is inspired by A. Hussain et al. [7] and Y. Wang
and drift in reading. Sensor faults can also be interpreted as et al. [8]. The SIM algorithm proposed by A. Hussain et al.
abnormal increase in stator current due to excessive overloads [7] can only detect a sensor fault of certain magnitude. Smaller
or fluctuations in speed, considering stator current and speed as sensor faults are not detectable using that algorithm. Detection
measurements 3) Actuator faults that occur due to any problem of these smaller sensor faults becomes important if this sensor
in 3 phase power supply which results into improper stator is used to control the parameters of a critical and sensitive
voltage and may ultimately lead to stator faults as described process, for example, a flow control valve adjusting the flow
previously. of coolant to reactor core in a nuclear power plant. This valve
Researchers have developed techniques to diagnose these is controlled by a level sensor measuring the coolant level in
different failures of induction motor using model based and core assembly. If the results of these level sensors are least
data-driven Fault Diagnosis (FD) schemes. As described in accurate then flow of coolant to the reactor core may badly
[1], fault modeling of induction motor was performed such affected resulting in excessive heating or damaging of core.
that, faults were detected in induction motor during its normal SIM based fault isolation algorithm of Y. Wang et al. (2011)
operating condition automatically. In [2], a model based fault does not consider the effect of process disturbance and sensor
detection and isolation strategy for induction motor was pre- noise on residual while computing the robust parity vectors. It
sented. It was based on non-linear observer in collaboration only increases the fault sensitivity to raise the impact of fault
with neural network which estimated the states of the system. on residual.
Any difference between estimated states and actual system In our work, an improved fault detection algorithm will be
states was considered as fault. In [3], modeling of induction proposed which detects smaller sensor faults. This is achieved
motor based on fuzzy logic was performed for fault diagnosis. by performance index ratio given in this work regarding SIM
Inputs to this fuzzy system model were stator current and time. algorithm. This index signifies the impact of small magnitude
This fuzzy modeling used some decision logics according sensor faults on residual signal. We will also propose a fault
to which the fault was detected. The above described FD isolation algorithm which considers the effect of both, the dis-
solutions are based on modeling of induction motor. Modeling turbance and fault on residual signal unlike previously existing
of induction machine is a difficult task especially when motor literature. For that purpose, isolation indices are modified and
is connected to other auxiliaries such as converters and other improved isolation algorithm is given in this work. Further,
electronic circuits. This pre-modeling may result into uncertain we will present the way of reducing False Alarm Rate (FAR),
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Fig. 1. Functional block diagram of three phase induction motor

which is usually caused by higher impact of disturbance and


noise during healthy conditions of the process. This reduction
in FAR is achieved by signal processing technique such as
wavelet transform. This reduced FAR will improve reliability
of fault diagnosis scheme.
Remaining of the paper is organized as follows: Section Fig. 2. Experimental setup of three phase induction motor
II includes three phase induction motor details, Section III is
related to foundation and preliminaries of data driven scheme,
Section IV presents improvement in performance index, Sec-
tion V deals with detailed fault isolation study, Section VI 
k1 0 k2 0
 
0 −p 0 0

includes experimental results along-with their analysis and  0 k1 0 k2   p 0 0 0 
finally the Section VII is conclusion. A0 =  , Aω =  ,
 k3 0 −γ 0   0 k4 0 0 
0 k3 0 −γ −k4 0 0 0
II. D ESCRIPTION OF T HREE PHASE I NDUCTION MOTOR
 
Three phase induction motor working is based on induction 0 0  
of stator voltage to its rotor. Hence, input applied is stator  0 0  0 0 1 0
B=
 k5
, C= , D = N ull(2 × 2)
voltage and is considered as actuator and measurement is stator 0  0 0 0 1
current which is proportional to torque produced. Functional 0 k5
block diagram of motor is given in Fig. 1. Here, ‘ud ’ and ‘uq ’,
 T
the stator voltages of d and q axes are actuators and ‘id ’ and Where, x = φq φd iq id is state vector, u =
‘iq ’, the stator currents of d and q axes are sensors. As shown  T  T
uq ud is input vector and y = iq id is output
by Fig. 1, four types of faults are possible f1 , f2 , f3 and f4 . vector. Test system for this research work is Three Phase
Modelling details of squirrel-caged rotor induction motor to Induction Motor, present in Electric Machines Lab, Pak-
obtain input-output data, are described below [9]: istan Institute of Engineering and Applied Sciences (PIEAS).
Schematic diagram is shown in Fig. 2. This setup constitutes
φ˙q = k1 φq + k2 iq − ωpφd
of a servo-machine at left in Fig. 2 coupled to a squirrel-caged
φ˙d = k1 φd + k2 id + ωpφq induction motor. Technical specifications of motor are given
(1)
i˙q = k3 φq − γiq + ωk4 φd + k5 uq below [10]:
i˙d = k3 φd − γid − ωk4 φq + k5 ud • Nominal line-line RMS voltage= 400V
• Frequency= 50Hz
• Nominal (full load) line current= 1A
1 Lsr Lsr pLsr • Synchronous speed= 3000rpm
k1 = − , k2 = , k3 = , k4 = ,
Trot Trot Trot σLs Lr σLs Lr • Nominal speed= 2800rpm
• Starting current to nominal current ratio= 7
1 Lr Rs Lsr 2 Lsr 2 • Nominal power factor= 83%
k5 = , Trot = ,γ = + , σ = 1−
σLs Rr Ls σ Trot σLs Lr Ls Lr It is assumed that, motor is operating at nominal speed of
Here, ‘φd ’ and ‘φq ’ are the rotor fluxes of d and q axes. ω = 293rad/s. Parameters of this squirrel-caged induction
Remaining parameters are described in Table I. In generic form motor are derived based on the specifications described pre-
we can write: viously and these are shown in Table I. Discretized Model of
motor is given below in (3):

ẋ = (A0 + ωAω )x + Bu x(k + 1) = Ad x(k) + Bd u(k)


(2) (3)
y = Cx + Du y(k) = Cd x(k) + Dd u(k)
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

TABLE I
N OMINAL PARAMETERS OF I NDUCTION M OTOR [10]    
0 0
... 0 D 0 ... 0
 C 0
... 0   CB D ... 0 
Gi f =  , G d
=
   
Symbol Name Value .. ...
.. .  i
  .. .. . . .. 
 . ..
.  . . . .
Ls Stator inductance 0.1164H CAs−1 CAs−2 ... 0 CAs−1 B .. ... D
(7)
Lr Rotor inductance 0.1164H
SIM based data-driven method is governed by the following
Lsr Mutual inductance 0.1109H equation [11], [12]:
σ Flux leakage factor 0.092896
Rs Stator resistance 0.7941ohm Yf = Γs Xi + Gi d Uf + Gi f Pf + Sf (8)
Rr Rotor resistance 3.5735ohm Equation (8) is a modified form of (5) such that, (8) is in data-
p Number of poles 2 driven and (5) is in model based framework. Here, Yf and Uf
are future output and input block hankel matrices respectively,
Γs is observability matrix, Xi is state matrix, Pf and Sf are

−0.3091 −0.9001 0.004102 −0.005344
 process and sensor noise block hankel matrices respectively.
 0.9001 −0.3091 0.005344 0.004102  Past and future input block hankel matrices are given by:
Ad = 
 84.37

58.01 0.1985 0.3444  
Up = us (k) us (k + 1) ... us (k + N − 1)

−58.01 84.37 −0.3444 0.1985  
Uf = us (k + s + 1) us (k + s + 2) ... us (k + s + N )
 
42.18 29 0.5992 0.1722
Cd =
   
−29 42.18 −0.1722 0.5992 u0 u1
... uj−1 ui ui+1 ... ui+j−1
 u1 u2
... uj   ui+1 ui+2 ... ui+j 
Up =  , Uf =  .
   
.. . ..
. .
.  . .. . . .. 
   . .. .   . . . . 
0.001897 −0.002471
 0.002471 0.001897 
  ui−1 ui−2 ... ui+j−2 u2i−1 u2i ... u2i+j−2
0.2771 0.07963
Bd =   , Dd =
 0.5542 0.1593  −0.07963 0.2771 Here, Up ∈ Rsp∗l×N , Uf ∈ Rsf ∗l×N , Yp ∈ Rsp∗m×N and
−0.1593 0.5542 Yf ∈ Rsf ∗m×N . Where, s = i − 1 and (sp , sf ' s) > n.
The sampling period of 0.01sec is taken for discretization of Similarly, past and future output block hankel matrices denoted
the model in (2). It is worth noting that this discrete model of by Yp and Yf can be constructed. Some other required matrices
 T
asynchronous motor is used for input-output data generation. denoted by Zp and Zf are formed as Zp = Yp Up and
 T
Zf = Yf Uf . Equation (8) can be written into another
III. DATA -D RIVEN FAULT DETECTION S CHEME way:

Γs Gi d Gi f Pf + Sf
    
Consider a discrete linear time invariant system given as Xi
Zf = + ψ, ψ =
follows: 0 I Uf 0
x(k + 1) = Ax(k) + B(u(k) + fa (k)) + w(k) By pre-multiplying both sides of above equation with (1/N )
(4)
y(k) = Cx(k) + D(u(k) + fa (k)) + v(k) + fs (k) and post-multiplying with Zp T gives:

Γs Gi d
  
Here, A ∈ Rn×n , B ∈ Rn×l , C ∈ Rm×n and D ∈ Rm×l . 1 T 1 Xi 1
Zf Zp = Zp T + ψZp T
Where, x(k) ∈ Rn×1 is a state vector, u(k) ∈ Rl×1 is the N N 0 I U f N
input vector. fa (k) is actuator fault vector, fs (k) is sensor
fault vector, w(k) is process noise or disturbance and v(k) For N → ∞ we have (1/N )ψZp T = 0. Thus, doing
is sensor noise. By using (4), we can write expressions for Singular Value Decomposition (SVD) of covariance matrix
y(k − s), y(k − s + 1) and y(k) recursively and combining all ‘( N1 Zf Zp T )’ will give us parity space ‘Γs ⊥ ’ (Γs ⊥ Γs = 0)
these expressions we obtain the following: and corresponding ‘Γs ⊥ Hus ’, as described below in algorithm
1, where Hus ∼ Gi d .
ys (k) = Γs x(k − s) + Gi d us (k) + Gi d fa,s (k)
(5)
+ Gi f ws (k) + vs (k) + fs,s (k) Algorithm 1
 T • Estimate the system order ‘n’
ys (k) = y(k − s) y(k − s + 1) ... y(k) 1 T
• Do SVD of covariance matrix: N Zf Zp = Uz Σz Vz
 
us (k) =

u(k − s) u(k − s + 1) ... u(k)
T Uz11 Uz12
Uz = , Uz11 ∈ R(msf )×(lsf +n) ,
Uz21 Uz22
Similarly, the other vectors fa,s (k),fs,s (k),vs (k) and ws (k) Uz12 ∈ R(msf )×(msf −n) , Uz22 ∈ R(lsf )×(msf −n)
can be constructed.
• Extract the terms Γs ⊥ Hus and Γs ⊥ using Γs ⊥ = Uz12 T
and Γs ⊥ Hus = −Uz22 T
 s
T
Γs = C CA ... CA (6)
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

T
IV. P ERFORMANCE INDEX TO ACHIEVE HIGHER solving J1 in (9) means that the term Γs ⊥ Hus Hus T Γs ⊥ is
SENSITIVITY TO SENSOR FAULTS maximized, due to which, influence of term “Γs ⊥ Hus us (k)”
Once, the terms Γs ⊥ Hus and Γs ⊥ are obtained from input- signifies in (11) and may become nearly equal to the term
output data of system using algorithm 1, then next goal is to “αrob ys (k)”. Consequently, a very small change or no change
compute optimal parity vector to achieve robustness against occurs in residual signal for these smaller sensor faults. So, if
T
disturbances and sensitivity to faults. To achieve these two we use the index of (12) then, the term (Γs ⊥ Hus Hus T Γs ⊥ )
goals simultaneously, A. Hussain et al. (2016) proposed the is minimized and influence of term “Γs ⊥ Hus us (k)” in (11)
following performance index. is reduced. Thus, even for smaller sensor faults we get a
T
higher change in residual signal. Although, J2 causes higher
Γs ⊥ Hds Hds T Γs ⊥ sensitivity to sensor faults but the upper bound is given
J1 = min T
(9)
Γs ⊥ Γs ⊥ Hus Hus T Γs ⊥ by (16):
||Γs ⊥ Hus us (k)|| > ||us (k)||− (16)
where Hus ∼ Gi d and Hds ∼ Gi f in (8). Solution of (9) is
found by considering it as eigen-value problem: Above expression describes that, J2 should not minimize
Γs ⊥ Hus us (k) the term so much that, lower magnitude
T T
ls,min (Γs ⊥ Hds Hds T Γs ⊥ − λs,min Γs ⊥ Hus Hus T Γs ⊥ ) = 0 actuator faults become difficult to detect. There is an-
(10) other side of performance index J2 that, it is maximizing
T
Where, ls,min is minimum eigen-vector and λs,min is mini- Γs ⊥ Hds Hds T Γs ⊥ which causes to increase the influence of
mum eigen-value. After identification of terms Γs ⊥ Hus and disturbance on residual but it will not disturb the sensitivity
Γs ⊥ , matrix Hds is also required for solving (10). So, to to faults. Its sole consequence is the increase in FAR which
construct Hds matrix, A and C matrices are extracted from can be reduced by the post-processing of residual which will
Γs as shown in (6), whereas, Γs = null(Γs ⊥ ). Robust parity be presented later in this work. As far as, actuator faults
vector based residual generator is given as follows [7]: are concerned, then performance index J2 also detects them
successfully. Its justification is that, during actuator fault,
Rrob (k) = αrob ys (k) − Γs ⊥ Hus us (k) (11)
the term “Γs ⊥ Hus us (k)” in (11) dominates due to abnormal
Here, αrob = ls Γ⊥ is optimal parity vector corresponding to increase in us (k), thus, causing a higher change in residual.
index J1 in (9). As shown in (7), Hus ∼ Gi d is showing the This completes the proof.
effect of actuators via B and D matrices. Similarly, the effect Solution of J2 is also given by (10) with only difference that
of sensors is represented by Hds ∼ Gi f in (7). Hence, the now ‘ls,min ’ is replaced by maximum eigen-vector ‘ls,max ’
index J1 in (9) only increases the effect of actuator faults on and ‘λs,min ’ by maximum eigen-value ‘λs,max ’. Performance
residual and reduces the impact of disturbance on residual. index in (12) is justified because Hds ∼ Gi f is showing the
On other hand, it results into a fault detection scheme which effect of sensors or measurements on residual and Hus ∼ Gi d
is lesser sensitive to sensor faults. We propose an improved is showing the effect of inputs on residual. Whereas, these
version of the index (9) in the following lemma. inputs may be system input or noise.

Lemma 1: To achieve higher sensitivity to sensor faults, V. R OBUST S ENSOR AND ACTUATOR FAULT ISOLATION
performance index given below in (12) should be used A. Fault Isolation
T Fault isolation means to determine which sensor or actuator
Γs ⊥ Hds Hds T Γs ⊥
J2 = max T
(12) gets faulty. Its algorithm is based on finding parity space for
Γs ⊥ Γs ⊥ Hus Hus T Γs ⊥ i-th sensor such that its effect on i-th residual is completely
P roof : Parity space based residual generator is given by [14]: decoupled from the effect of remaining sensor. Fault isolation
algorithm is presented below [8]:
r(k) = α(ys (k) − Hus us (k)) (13)
Here, ‘α’ is a parity vector and ‘r(k)’ denotes residual signal. Algorithm 2
⊥ ⊥
Substitution of (5) into (13) and using (αΓ = 0) gives us: • Apply Algorithm 1 to obtain Γs Hus and Γs from
input-output data of process
r(k) = α(Hus fa,s (k) + Hds ws (k) + vs (k) + fs,s ) (14) Sensor Fault Isolation:

Equation (14) in general form is given by (15): • To distribute Γs as follows:
⊥ ⊥
Γs,2 ⊥ ... Γs,sa ⊥ , where Γs,k ⊥ ∈
 
Γs = Γs,1
r = Hds Ws + Hus fa + fs (15) Rµ×m , k = 1, 2, ..., sa and µ = msf − n

In case of only small magnitude sensor fault (fs → 0) and • Now sub-divide Γs,k as below:
⊥ ⊥
Γs,k,2 ⊥ ... Γs,k,m ⊥
 
no actuator fault (fa = 0), (15) reduces to r ≈ Hds Ws . Γs,k = Γs,k,1 where
This shows that in case of smaller sensor faults, residual has Γs,k,i ⊥ ∈ Rµ×1 , i = 1, 2, ..., m

nominal value caused by the noise or we can say there is • Now re-arrange Γs,k,i , i = 1, 2, ..., m in following

no alternation in residual. Another way to prove this is: if way to decouple the effect of other sensors: Γs,k,ĩ ⊥ =
there is a sensor fault of very lower magnitude, it implies [ Γs,1,1 ⊥ ... Γs,1,i−1 ⊥ Γs,1,i+1 ⊥ ... Γs,1,m ⊥ ...
that the term “αrob ys (k)” in (11) remains very small. Then Γs,sa ,1 ⊥ ... Γs,sa ,i−1 ⊥ Γs,sa ,i+1 ⊥ ... Γs,sa ,m ⊥ ]
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Solve Ssen,i Γs,k,ĩ ⊥ = 0 to obtain Ssen,i design nmnf−1 number of residuals [14] and a table is con-


⊥ structed between faults and designed residuals. From this table
• Find parity space for i-th sensor ‘Γs,sen,i ’ using
⊥ ⊥ by observing the sequences, we can select fewer number of
Γs,sen,i = Ssen,i Γs
Actuator Fault Isolation: residuals to detect all possible faults.

• To divide Γs Hus as:

Γs Hus = Γs,1 ⊥ Hus Γs,2 ⊥ Hus ... Γs,sa ⊥ Hus , B. Residual Evaluation or Statistical Testing
 

where Γs,k ⊥ Hus ∈ Rµ×l For the sake of completion of Fault Detection process,
⊥ statistical testing of residual is performed using Generalized
• Sub-divide Γs,k Hus as:
Γs,k,1 Hus Γs,k,2 ⊥ Hus ... Γs,k,l ⊥ Hus

 
where Likelihood Ratio (GLR) algorithm presented in [15], [16]. As
⊥ µ×1 it is obvious, that the residual is affected by noise and fault,
Γs,k,i Hus ∈ R , i = 1, 2, ..., l

• Re-arrange Γs,k,i Hus , i = 1, 2, ..., l as below for so we consider this noise as white noise N (0, σ 2 ). Major
decoupling the influence of other actuators: steps of GLR algorithm are given below:
Γs,k,ĩ ⊥ Hus = [ Γs,1,1 ⊥ Hus ... Γs,1,i−1 ⊥ Hus
Γs,1,i+1 ⊥ Hus ... Γs,1,l ⊥ Hus ... Γs,sa ,1 ⊥ Hus Algorithm 3
... Γs,sa ,i−1 ⊥ Hus Γs,sa ,i+1 ⊥ Hus ... Γs,sa ,l ⊥ Hus ] • From Chi-square distribution table, find χa such that

• Solve Sact,i (Γs,k,ĩ Hus ) = 0 to obtain Sact,i
prob[χ > χa ] = α to set allowed limit of FAR
χ
⊥ ⊥ ⊥ • To set threshold as Jth = 2a
• Find ‘Γs,act,i ’ using Γs,act,i = Sact,i Γs PNw
1 2
• Testing statistics are given by J = 2σ 2 N ( i=1 r(i) )
Next objective is to achieve higher sensitivity to faults w
• Decide as:
and robustness against disturbances so that the results of
J < Jth =⇒ No fault, J > Jth =⇒ Fault occurs
fault isolation algorithm are reliable. This is achieved by our
proposed performance indices for fault isolation which are
supported by Lemma 1 described previously. These indices C. Post-processing of Residual
are given in (17) below: Due to higher impact of disturbances on residual, there may
be significant FAR. Effect of noise on residual appears as
T higher frequency content. To remove it from residual, wavelet
Γs,sen,i ⊥ Hus Hus T Γs,sen,i ⊥
Jsen,i = min T
transformation is used [17]. Wavelet transform decomposes the
Γs ,sen,i ⊥ Γs,sen,i ⊥ Hds Hds T Γs,sen,i ⊥ residual into approximation and detailed coefficients. Detailed
(17)
T coefficients contain higher frequency content (noise) as well
Γs,act,i ⊥ Hus Hus T Γs,act,i ⊥
Jact,i = max T
as the information about the change in residual caused by
Γs ,act,i ⊥ Γs,act,i ⊥ Hds Hds T Γs,act,i ⊥ fault. So, motivated by [17], we used wavelet transformation
Our proposed indices in (17) are better than that in [8] to reduce this higher frequency content. During this signal
because performance index in [8] only increases the effect processing, noise is removed by suppressing higher frequency
of fault on residual signal while, not considering the noise content as well as retaining the change in residual signal
affecting the system. Now, for our proposed indices in (17), caused by fault. However, after wavelet transformation,
we require Hds ∼ Gi f matrix, which is constructed as shown there exists some FAR caused by lower frequency process
by (7). So, we extract A and C matrices from observability disturbances. This FAR can be minimized by high pass
matrix ‘Γs ’, which is shown in (6). Where, Γs is obtained filtering of wavelet processed residual signal.
using (Γs ⊥ Γs = 0).
VI. R ESULTS AND DISCUSSION
• To evaluate robust parity vectors as:
The input output data of the induction motor is obtained by
αsen,i = lmin,i Ssen,i Γs ⊥ , αact,i = lmax,i Sact,i Γs ⊥ running the motor for 90 seconds. Since, the data is noisy so it
Here, lmin,i and lmax,i are obtained by solving indices is assumed that the process noise and measurement noise are
given in (17) white guassian with power spectral densities of 0.2 and 0.01
• Build residuals for i-th sensor and actuator using follow- respectively. Total 9992 samples are recorded. There are two
ing equations: sensors id and iq , which means m = 2. Similarly, there are
two actuators ud and uq and l = 2. Values of sp and sf are
rsen,i (k) = αsen,i (ys (k) − Hus us (k)) set to sp = sf = 5 higher than system order n = 4. Hence,
(18)
ract,i (k) = αact,i (ys (k) − Hus us (k)) we have Yp , Yf ∈ R10×9992 and Up , Uf ∈ R10×9992 . After
• Post-processing of each of the residuals to reduce FAR applying algorithm 1, Γs ⊥ Hus ∈ R6×10 and Γs ⊥ ∈ R6×10
are obtained.
Up till now, it was considered that there is a separate
residual for each of the sensors and actuators. It means,
for ‘j’ number of sensors and ‘p’ number of actuators, we A. Improved Fault Sensitivity to Lower Magnitude Sensor
require ‘nm = j + p’ number of residuals in order to isolate Faults
‘nf = j+p’ faults. Sometimes, due to complexities of process, A very small sensor fault of 0.5A in q-axis current ‘iq ’ is
it is economical to detect ‘nf ’ faults using measurements applied. This sensor fault is caused by insulation failure of
lesser than ‘nf ’ which means that (nf > nm ). Then, we stator winding and as a result, short circuits in stator winding
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Residual Signal for Performance Index "J1" Residual Signal for Performance Index "J1"
4 10

8
3
Residual

Residual
6
2
4

1
2

0 0
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Time (s) Time (s)

Residual Signal for Performance Index "J2" Residual Signal for Performance Index "J2"
4 10

8
3
Residual

Residual
6
2
4

1
2

0 0
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Time (s) Time (s)

Fig. 3. Comparison between J1 and J2 for lower magnitude sensor fault Fig. 5. Comparison between J1 and J2 for lower magnitude actuator
fault
Residual Signal for Performance Index "J1"
20
Residual for sensor#1 Residual for sensor#2
20 20

residual for q-axis current (I q )

residual for d-axis current (I d )


15
Residual

15 15
10

10 10
5

5 5
0
0 10 20 30 40 50 60 70 80 90
Time (s) 0 0
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Residual Signal for Performance Index "J2" Time (s) Time (s)
20
J-statistics for sensor#1 J-statistics for sensor#2

J-Statistics for q-axis current (I q )

J-Statistics for d-axis current (I d )


3 3
15
2.5 2.5
Residual

10 2 2

1.5 1.5
5
1 1

0 0.5 0.5
0 10 20 30 40 50 60 70 80 90
Time (s) 0 0
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Time (s) Time (s)

Fig. 4. Comparison between J1 and J2 for higher magnitude sensor


fault Fig. 6. Sensor:1 is faulty and sensor:2 is healthy. Fault of sensor:1 is
isolated from sensor:2

occur. Ultimately, measured stator current is increased. Then,


separate residual for each of the sensors. After that, algorithm
J1 and J2 are solved using (9) and (12), to get ‘ls,min ’
3 was applied on these 2 residuals individually to obtain
and ‘ls,max ’ respectively. Then, two robust parity vectors
their GLR based J-statistics. While applying GLR algorithm,
‘αrob,min = ls,min Γs ⊥ ’ and ‘αrob,max = ls,max Γs ⊥ ’ are
Nw = 10 was selected. It means, we take the window of 10
obtained. Finally, two residual generators are constructed and
samples of residual, to calculate mean square value required
their comparison is shown in Fig. 3. It can be seen in Fig. 3
by J-statistics. Results are shown in Fig. 6. We can see that
that for smaller sensor fault there is no significant change in
only the residual of sensor:1 is changing and no change in
residual corresponding to J1 , whereas, residual corresponding
residual of sensor:2.
to J2 gives a noticeable change and this small sensor fault
Now, we observe actuator fault isolation. Consider the case
is detected. Then, the effect of larger magnitude sensor fault
when only actuator:1 is faulty and there is no fault on
is studied on stator current, which is caused by excessive
actuator:2, results after applying algorithms 2 and 3 are
overloading of motor and same process is repeated. Results
shown in Fig. 7. Only the residual of faulty actuator:1 is
for this case are shown in Fig. 4. It is obvious from results
changing. For J-statistics of actuators, Nw = 6 is selected.
that, smaller sensor faults can be successfully detected using
The difference between isolation of actuator and sensor faults
J2 . Now, a small actuator fault is applied such that, d-axis
lies in performance indices used for them which are shown
stator voltage ‘ud ’ deviates from normal value. Again, similar
in (17). To check the effectiveness of our proposed isolation
analysis is performed and results are shown in Fig. 5. It can
algorithm, we plot the residuals in fault free case for sensor:1
be observed that in case of smaller actuator fault, both J1 and
using algorithm of [8] and our proposed isolation algorithm.
J2 perform correctly.
Now, comparison of Fig. 8 and Fig. 9 shows that FAR is much
lower in our proposed algorithm. This shows that for a given
B. Fault Isolation process, results of our proposed isolation algorithm are more
Here, following notations will be used: reliable than that of [8] because of increased robustness against
• Sensor:1 → sensor measuring ‘iq ’ disturbances, as described previously in section V-A.
• Sensor:2 → sensor measuring ‘id ’
• Actuator:1 → q-axis stator voltage ‘uq ’ C. Results for Fault Isolation Using Fewer Number of
• Actuator:2 → d-axis stator voltage ‘ud ’ Sensors
First, we consider sensor fault isolation. Let us select the case Now, to detect these four types of faults i.e. 2 for sensors
when only sensor:1 is faulty and there is no fault on sensor and 2 for actuators, instead of 4 residuals we will consider
sensor:2. Now, fault isolation algorithm 2 is applied. There is only 3 residuals. Two residuals ‘R1’ and ‘R2’ will deal sensor
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

1.5
Residual for actuator#1
1.5
Residual for actuator#2 TABLE II
residual for q-axis voltage (V q )

residual for d-axis voltage (V d )


FAULT ISOLATION LOGIC TABLE
1 1

0.5 0.5
Fault R1 R2 R3
0 0
0 10 20 30 40
Time (s)
50 60 70 80 90 0 10 20 30 40
Time (s)
50 60 70 80 90
f1 (Sensor:1 faulty) 1 0 1
J-statistics for actuator#1 J-statistics for actuator#2
f2 (Sensor:2 faulty) 0 1 1
J-Statistics for q-axis voltage (V q )

J-Statistics for d-axis voltage (V d )


4 4

3 3
f3 (Actuator:1 faulty) 1 1 1
2 2
f4 (Actuator:2 faulty) 1 1 0
1 1

0 0
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Time (s) Time (s)

residual for I q (Ampere)


R1
100

50
Fig. 7. Actuator:1 is faulty and actuator:2 is healthy. Fault of actuator:1
is isolated from actuator:2 0
0 10 20 30 40 50 60 70 80 90

residual for I d (Ampere)


R2
30

20
residual for sensor#1 using Old algorithm
14
10

0
12 0 10 20 30 40 50 60 70 80 90

R3

residual for V q (Volts)


4

10
residual for q-axis current (I q )

8
0
0 10 20 30 40 50 60 70 80 90
time (s)
6

Fig. 10. Verification of Table II, for example, fault ‘f4 ’ (Actutor:2 is faulty)
2 is detected by combination “110”
0
0 10 20 30 40 50 60 70 80 90
time (s)
TABLE III
FAR AND FDR IN FOR EACH SENSOR AND ACTUATOR FAULT
Fig. 8. Residual in fault-free case, showing that the effect of noise on
residual is higher, plotted using algorithm in [8] f1 f2 f3 f4

residual for sensor#1 using proposed algorithm


FAR 3.43 2.67 0.1 0.3
14
FDR 96.28 97.14 98.94 98.06
12

10
residual for q-axis current (I q )

8
the the principle of Stein’s Unbiased Risk [20]. Then, soft
thresholding is performed on detailed coefficients. Since, noise
6

contribution in detailed coefficients has amplitude lower than


4
the threshold selected, so it will be suppressed. But fault has
2 amplitude much higher than threshold, so its information will
0
be retained. Results for the case when actuator:2 is faulty
0 10 20 30 40 50 60 70 80 90
Time (s)
and no fault on actuator:1, are shown in Fig. 11. It can be
observed that due to application of wavelet transform, FAR is
Fig. 9. Residual in fault-free case, showing that the effect of noise on significantly reduced as shown by comparison of original and
residual is lower, plotted using our proposed isolation algorithm
de-noised J-statistics of actuator:1. Table III shows FAR and
FDR in each of the four types of faults, while using algorithms
fault isolation and one ‘R3’ relates to actuator fault isolation. proposed in this work.
Using the scheme presented at the end of section V-A, we
construct fault isolation logic table referred as Table II. It tells VII. O NLINE I MPLEMENTATION OF PROPOSED SIM
BASED FAULT DIAGNOSIS
that, fault f1 is detected if residuals ‘R1’ and ‘R3’ are active
and residual ‘R2’ is not active. Results verifying Table II are During practical implementation of FD technique pre-
shown in Fig. 10. sented in this work, it is required to take input-output
data samples of system under observation. Then these data
samples are fed to Fault Detection and Monitoring (FDM)
D. Post-processing of residual signal system for evaluation of system as healthy or faulty. During
To further reduce the impact of disturbances on residual in evaluation of system, the system can change its state from
order to reduce FAR, wavelet de-noising technique [18], [19] is healthy to faulty e.g. large amplitude transients or surges
applied. First, we select ‘db4’ as mother wavelet and 3 number (of few m-sec) in 3 phase supply to induction motor. The
of decomposition levels. After that, wavelet decomposition of chances of missing these faults can be reduced by Recursive
residual gives us detailed and approximate coefficients. To Subspace Identification Method along with a predictive
reduce the impact of noise, threshold is selected based on control in [21].
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

4
Original J-Statistics of actuator#1
4
Original J-Statistics of actuator#2 [2] R. S. Kraleti, M. Zawodniok, and S. Jagannathan, “Model based
diagnostics and prognostics of three-phase induction motor for vapor
3 3
J-Statistics
compressor applications” in Proc. IEEE Conference on Prognostics and

J-Statistics
2 2
Health Management, DOI 10.1109/ICPHM.2012.6299525, pp. 1-7, Sep.
1 1 2012.
0
0 10 20 30 40 50 60 70 80 90
0
0 10 20 30 40 50 60 70 80 90
[3] S. O. Ibrahima, k. N. Faris and E. A. Elzahab, “Implementation of
Time (s) Time (s) fuzzy modeling system for faults detection and diagnosis in three phase
De-Noised J-Statistics of actuator#1 De-Noised J-Statistics of actuator#2
4 4 induction motor drive system”, Jour. of Elect. Sys. and Info. Tech., Vol.
3 3 2, No. 1, pp. 27-46, May 2015.
J-Statistics

J-Statistics
2 2
[4] F. Ferracuti, A. Giantomassi, S. Iarlori, G. Ippoliti and S. Longhi, “In-
1 1 duction Motor Fault Detection and Diagnosis using KDE and Kullback-
0 0
Leibler Divergence”, Industrial Electronics Society, IECON Annual
0 10 20 30 40 50 60 70 80 90 0 10 20 30 40 50 60 70 80 90
Time (s) Time (s) Conference of IEEE, pp. 2923-2928, 2013.
[5] R. Islam, S. Ali Khan and J.M. Kim, “Discriminant Feature Distribution
Analysis-Based Hybrid Feature Selection for Online Bearing Fault
Fig. 11. Wavelet de-noising when Actuator:2 is faulty, FAR is reduced Diagnosis in Induction Motors”, Journal of Sensors, Hindaw, vol. 2016,
in de-noised J-statistics of actuator:1 and no effect on J-statistics of DOI 10.1109/IECON.2013.6699595, pp. 1–16, Jan. 2016.
actuator:2 [6] S. X. Ding, P. Zhang, A. Naik, E. Ding, and B. Huang, “Subspace
method aided data-driven design of fault detection and isolation sys-
tems,” Jour. of Process Control, Vol. 19, No. 9, pp. 1496–1510, 2009.
VIII. CONCLUSION [7] A. Hussain, A. Q. Khan and M. Abid, “Robust fault detection using
subspace aided data driven design”, Asian Jour. of Control, Vol. 18,
DOI 10.1002/asjc.1086, No. 2, pp. 1–12, Mar. 2016.
In this work, SIM based data driven approach was used [8] Y. Wang, G. Ma, S.X. Ding and C. Li, “Subspace aided data-driven
for Fault Detection and Isolation (FDI) of induction motor. design of robust fault detection and isolation systems”, Automatica, Vol.
Smaller sensor faults are successfully detected using our 47, DOI 10.1016/j.automatica.2011.05.028, No. 11, pp. 2474–2480, Nov.
2011.
proposed performance index regarding SIM based fault detec- [9] Y. Rachedi, M. O. Rachedi and S. Yahmedi, “Robust control design
tion. Moreover, sensor and actuator fault isolation algorithm of an induction motor”, Int. Jour. of Electrical Energy, Vol. 1, DOI
presented in this paper is more robust against undesired signals 10.12720/ijoee.1.4.201-205, No. 4, pp. 201–205, Dec. 2013.
[10] User’s manual, Asynchronous machine 300W, DEE Machines Lab, Pak-
affecting the system. In a huge industrial process, large pumps istan Institute of Engineering and Applied Sciences (PIEAS), Islamabad,
are used for circulation of some essential fluids e.g. coolants Pakistan.
and lubricants. So this FDI scheme is helpful there to diagnose [11] S. Yin, Member, IEEE, S. X. Ding, X. Xie, Student Member, IEEE,
and H. Luo, “A review on basic data-driven approaches for industrial
the faults of motor as quickly as possible. Along-with, it is process monitoring”, IEEE Trans. on Ind. Electronics, Vol. 61, DOI
desired that results of FDI scheme should be accurate whereas, 10.1109/TIE.2014.2301773, No. 11, Nov. 2014.
the accuracy of FDI scheme is badly affected by noise or un- [12] S. Yin , S. X. Ding , A. H. Abandan Sari and H. Hao, “Data-driven
monitoring for stochastic systems and its application on batch process”,
known inputs to the process. So our proposed FDI scheme Int. Jour. of Syst. Sci., Vol. 44, DOI https://doi.org/10.1080/00207721.
fulfills these requirements by achieving ‘robustness against 2012.659708, No. 7, pp. 1366–1376, Feb. 2013.
undesired signals’ and ‘higher sensitivity to faults’. In this [13] M. Green and David J.N. Limebeer, “Linear robust control”, Pearson
Education, Inc., 1995.
work, SIM based data driven approach was used for Fault [14] S.X. Ding, “Model-based fault diagnosis techniques: design schemes,
Detection and Isolation (FDI) of induction motor. A fault algorithms, and tools”, Springer, London, 2008.
detection algorithm based on improved performance index is [15] S.X. Ding, P. Zhang, A. Naik, E.L. Ding, and B. Huang, “Subspace
method aided data-driven design of fault detection and isolation sys-
proposed for successful detection of smaller sensor faults. To tems”, Jour. Process Control, Vol. 19, No. 9, pp. 1496-1510, 2009.
find location of a fault, a fault isolation algorithm is proposed [16] S. X. Ding, “Data-driven Design of Fault Diagnosis and Fault-tolerant
which is robust against undesired signals. This robustness im- Control Systems”, Springer, (2014).
[17] A. R. Khan, A. Q. Khan, M. T. Raza, M. Abid and G. Mustafa, “Design
proves the results of fault isolation algorithm by reducing FAR. of Robust Fault Detection Scheme for Penicillin Fermentation Process,”
Residual signals generated via our proposed FDI algorithms IFAC, Vol. 48, DOI 10.1016/j.ifacol.2015.09.590, No. 21, pp. 589–594,
are evaluated using thresholds. These algorithms are applied 2015.
[18] D. L. Donoho, “De-noising by soft-thresholding”, IEEE Trans. on Inf.
to a 3 phase induction motor to demonstrate the effectiveness Theory, Vol. 41, No. 3, pp. 613–627. 1995.
of our proposed algorithms. [19] Albert C. To, J. R. Moore and S. D. Glaser, “Wavelet denoising tech-
Future Work in this research includes 1) Fault identification niques with applications to experimental geophysical data”, Int. Jour. of
Signal Processing, Elsevier, Vol. 89, DOI 10.1016/j.sigpro.2008.07.023,
to obtain information about the size and nature of the fault 2) No. 2, pp. 144–160, Feb. 2009.
SIM based Fault Tolerant Control, so that system can retain [20] Y. Le Montagner, St. Member IEEE, E. D. Angelini, Senior Member
its normal behaviour during fault as long as possible. IEEE, and J-Christophe O-Marin, Fellow IEEE, “An unbiased risk
estimator for image denoising in the presence of mixed poisson gaussian
noise”, IEEE Tran. on Image Proc., Vol. 23, DOI 10.1109/TIP.2014.
2300821, No. 3, Mar. 2014.
R EFERENCES [21] P. Zhou, P. Dai, H. Song and T. Chai, “Data-driven recursive subspace
identification based online modelling for prediction and control of
[1] A. Selvanayakam, W.R. Babu and S.K. Rajarathna, “Fault detection in molten iron quality in blast furnace ironmaking,” IET Control Theory
three phase induction motor”, Int. Jour. of Innovatiove Reaserch in Sci. and Applications, Vol. 11, DOI 10.1049/iet-cta.2016.1474, No. 14, pp.
Engg. and Tech., Vol. 4, No. 4, pp. 84–90, Apr. 2015. 2343–2351, Sep. 2017.

You might also like