Condition Monitoring and Fault Diagnosis of

Generators in Power Networks

Bowen Yang, Fangyu Li, Jin Ye, Wenzhan Song
Center for Cyber-Physical Systems
University of Georgia
Athens, Georgia USA
Email: {bowen.yang, fangyu.li, jin.ye, wsong} @ uga.edu

Abstract—In this paper, a novel hierarchical signal processing In the past decade, many condition monitoring and fault
methodology is proposed for generator condition monitoring diagnosis methods for generators, including signal-based,
and fault diagnosis based on raw electrical waveform data in model-based and data-driven, have been investigated [1]–
power networks, which can often be measured by strategically-
located waveform sensors. The impact of generator short circuit [4]. Motor current signature analysis (MCSA) based on the
faults on strategically located electrical waveform sensors in frequency analysis of the stator currents has been one of the
power networks are firstly investigated and validated in Matlab most popular noninvasive condition monitoring and diagnosis
Simulink. Based on the large set of electrical waveform data methods. MCSA technique is mostly used to identify both rotor
produced by Matlab Simulink, a hierarchical algorithm is then faults (e.g., bearing damage, broken rotor bar, eccentricity,
designed to locate fault site location and monitor the condition of
generators in power networks. Finally, the proposed methodology end ring breakage, etc) and stator faults (e.g., short circuit
is validated in 14-bus IEEE standard power network under fault, etc) based on slots related harmonics [5], [6], third
different scenarios (e.g, one generator fault, two-generator-fault, harmonic [7], the sideband frequency components [8], and
various aging levels, etc). Our results show that we can locate the other [9]. The frequency analysis of stator voltages is
fault site location and monitor the aging condition of generators also used to detect stator winding inter-turn short faults in
in power networks. Compared to traditional condition monitoring
and fault diagnosis based on generator sensors, our proposed some operating conditions [10], [11]. Symmetrical component
methodology can monitor a large number of generators based analysis, which decomposes stator current or voltage to positive-
on a limited number of waveform sensors, which promises to sequence, negative-sequence, and zero-sequence components, is
reduce the cost of the maintenance and improve the reliability one of the alternative condition monitoring and fault diagnosis
of the power grid. methods due to symmetry of stator windings under the healthy
Index Terms—Synchronous Generators, Power Networks, Con-
dition Monitoring, Fault Diagnosis, Short Circuit Faults, Ad- condition [12], [13]. Therefore, condition monitoring and fault
vanced Signal Processing diagnosis techniques based on generator sensors (e.g, voltage,
current, vibration, etc) have been widely used and promise to
I. I NTRODUCTION reduce unscheduled downtime, and maintenance costs.
Generators are playing a vital role in electrical power When generators are connected in power networks, cur-
generation. Synchronous generators have been major means rent/voltage signature signals of faulty electric machines will
to generate electric power over a century. In recent years, propagate through the power networks [14]. The raw electrical
due to the increased penetration of wind energy, the number waveform and signals (e.g., voltage, current, harmonics, power
of asynchronous generators have risen rapidly. Condition factor, etc) in power networks will likely change, which contain
monitoring and fault diagnosis of large generators in power rich information about condition of generators. In [15], [16],
networks are gaining more interest since generator faults can some waveform information (magnitude or phase) were used
lead to a catastrophic failure and then outages if not detected to identify parameters of generators, however, they were not
in the early stage. Due to aging or severe operating conditions, yet used to monitor the aging condition of the generators for
generators are subject to many different types of faults including scheduled maintenance. In addition, for condition monitoring
stator faults, rotor electrical faults and rotor mechanical faults. and fault diagnosis of generators, raw electrical waveforms at
Among these faults, stator winding inter-turn short fault due to higher sampling rate besides magnitude and phase mentioned
the aging of winding insulation is the most dominant, which above might be needed as fault/deteriorating condition of
account for over 25% of faults in generators. When the inter- generators will produce unusual harmonics.
turn short fault is progressing, the condition of stator windings To the best of our knowledge, there are no existing works in
deteriorates and it can lead to catastrophic failures (e.g., phase- condition monitoring/fault diagnosis of generators by analyzing
to-ground short circuit). raw electrical waveforms in power networks, which can be
measured by strategically located electrical waveform sensors.
Our research is partially support by NSF-1725636, NSF-CNS-1066391,
NSF-CNS-0914371, NSF-CPS-1135814, NSF-CDI-1125165, and Southern In this paper, we propose to develop advanced signal
Company. processing methodology for generator condition monitoring

1
 Eq[0]
and fault diagnosis based on strategically located waveform ia = − cos(ωt + α0 )
xd
senors in power networks with the goal to reduce the cost 00
Eq0 Eq[0] t
of the maintenance and improve the reliability of the power − ( 00 − ) exp(− 00 ) cos(ωt + α0 )
xd xd Td
grid. Firstly, we build the the equivalent model of synchronous 0
Eq[0] Eq[0] t
generators under short circuit faults. Secondly, we analyze −( 0 − ) exp(− 0 ) cos(ωt + α0 )
the impact of generator short circuit faults on strategically xd xd Td
(1)
located electric waveform sensors in power networks. Thirdly, E 00 t
− d0 exp(− 00 ) sin(ωt + α0 )
we build 14-bus IEEE standard power network model in Matlab x00q Tq
Simulink and different scenarios (e.g, one generator fault, two- U[0] 1 1 t
+ ( + 00 ) exp(− ) cos(α0 − δ0 )
generator-fault, various aging levels, etc) to produce a large set 2 x00d xq Ta
of electric waveform data for condition monitoring and fault U[0] 1 1 t
+ ( − 00 ) exp(− ) cos(2ωt + α0 + δ0 )
diagnosis. Finally, based on the electrical waveform data in 2 x00d xq Ta
power networks, we will develop data-driven signal analysis Similarly, the rotor current can be obtained by:
approach to locate fault sources and estimate the aging levels if = if [0]
of generators in power networks.
xad xσD U[0] cos δ0 (xd − x0d )U[0] cos δ0 t
+[ − ] exp(− 00 )
(xf xD − x2ad )x00d xad x0d Td
II. ANALYSIS OF THREE - PHASE SHORT FAULT
(xd − x0d )U[0] cos δ0 t (2)
When the short circuit faults happens in synchronous gener- + exp(− 0 )
xad x0d Td
ators, the current could be multiple times of the rating current xad xσD U[0] t
of the generator, which could bring catastrophic damages to the − exp(− ) cos(ωt + δ0 )
(xf xD − x2ad )x00d Ta
power grid. In this section, the short-circuit fault of synchronous
generators is analyzed.
Assume the generator operates at speed ω, then the flux III. IMPACT OF GENERATOR FAULTS ON ELECTRIC
linkage will be: ψ = Ψ0 cos(α0 + ωt). And the fault happens WAVEFORMS IN POWER NETWORKS
at t = 0, the induced current will produce ∆ψ to maintain the To simplify the analysis of short circuit fault, the method of
initial flux linkage ψ0 , so ∆ψ = ψ0 − Ψ0 cos(α0 + ωt). symmetrical components is used in this section to analyze
According to the flux linkage analysis above, stator current the impact of generator short circuit faults on the power
includes two parts: one is the DC component iap generating networks. A according to symmetrical components, any types of
the initial flux linkage ψ0 ; the other one is the AC component asymmetrical three-phase phasor could be decomposed to three
i0 at the synchronous frequency (f = 2π ω
), which generates the symmetrical three-phase phasors. Take current for example, (3)
rotating field to offset the rotor exciting field. shows the relationship between currents in two coordinates.
For the salient pole machine, to compensate the difference Phase b and phase c currents could be derived by similar
in magnetic resistance of d- and q- axis, there will be one equations. 
I˙a(1) I˙a
   
additional AC component, the frequency of which is twice the 1 a a2
 I˙a(2)  = 1  1 a2 a  I˙b  (3)
synchronous frequency of the stator current. Therefore, three 3 1
types of magnetic field will be induced by the stator. And these I˙a(0) 1 1 I˙c
three magnetic fields will then change the flux linkage in the ◦ ◦
where a = ej120 , a2 = ej240 ; I˙a(1) , I˙a(2) , I˙a(0) are the pos-
rotor winding. Therefore, to maintain its flux linkage in the itive, negative and zero sequence of phase currents, respectively.
rotor, the rotor will have similar armature reaction and induce Then sequence impedance is introduced to describe relationship
three types of currents. Since the rotor is rotating at angular between voltage and current in symmetrical components coor-
speed ω, frequencies of these currents will be: ω, 0 and −ω. It dinate. Thevenin’s equivalent circuit for sequence impedance
should be noted that −ω means the field rotates in the opposite is shown in figure1. It should be noted that the asymmetrical
direction. components in sequence impedance are eliminated by adopting
The current components which will attenuate are refereed the asymmetrical voltage source and the transmission line
to as free current ∆i, while others are refereed to as forced reactances are modeled in Zf f . For generality, the equivalent
current i∞ . The classification is shown in Table I. circuit equation could be derived by:
TABLE I: Short Current Classification. Ėeq − Zf f (1) I˙f a(1) = U̇f a(1)
forced current i∞ free current ∆i 0 − Zf f (2) I˙f a(2) = U̇f a(2) (4)
stator i∞ ∆i0 = i0 − i∞ iap , i2ω
rotor if [0] ∆if a ∆if ω 0 − Zf f (0) I˙f a(0) = U̇f a(0)
Three more equations are needed to solve the equations
According to the synchronous generator model and constant- with 6 unknown variables, which are derived from the fault
linkage theorem, the short circuit current of generators with conditions. And common faults in power system include two
damping windings could be deducted by: phase short fault, two phase grounded short fault, etc. The
related fault conditions are listed in Table II.

2
 A. Event Detection
Before identifying the location of an event, we must first
become aware of the occurrence of such event. Thanks to
the data-driven time series anomaly detection techniques, the
presence of the event can be detected continuously. The changes
of the nodal voltages and branch currents can be expressed as:

Fig. 1: Equivalent circuit of each phase sequence ∆Vn = Vn (t) − Vn (t − w), ∆Inp = Inp (t) − Inp (t − w), (7)

where, w is the analysis window size, n and p denote two

TABLE II: Common Fault Condition in Power System. arbitrary neighboring nodes. If abnormal changes happen to
fault category fault condition ∆Vn and ∆Inp , which indicate the difference between the pre-
single phase short U̇f a = 0 I˙f b = 0 I˙f c = 0 and post-event, an event can be detected. Once the occurrence
two phases short I˙f a = 0 I˙f b + I˙f c = 0 U̇f b = U̇f c of an event is detected, the next step is to identify the types
two phases grounded I˙f a = 0 U̇f b = 0 U̇f c = 0
of the event and location of the root cause of the event.

Therefore, common asymmetric short faults could be solved B. Phase Unbalance Characterization
by (4) and the related fault equations listed in Table II. Other Simulating the generator aging faults, we could meet single,
bus voltage and branch current in the normal condition could two or even three phase issues, which means the short circuit as
also be solved similarly. However, this method requires the shown in Figure 2. The waveforms of Phases A,B, and C allow
detailed grid topology information and the accurate fault a relatively straightforward phase unbalance characterization
location, which may not be easy to acquire. In addition, based on direct comparisons of phase signal attributes.
to simulate multiple fault cases, this method is based on a Based on the IA, we define the current unbalance character-
strong assumption that the whole system is linear, which is ization functions Iα , Iβ , and Iγ as:
hardly applicable in the real power system. Therefore, in the
i,j∈{A,B,C}
following section, an advanced signal processing technique will 1 X
be proposed to locate the fault source for condition monitoring Inα = (AIni − AInj )2 . (8)
3
and fault diagnosis. i6=j

IV. FAULT I DENTIFICATION AND L OCATION Imax − Imin

In this section, we use the measured waveform data to Inβ = × 100%, (9)
identify and locate generator faults. Algorithm 1 shows how Imax
the whole process works. The waveforms of voltage and i,j∈{A,B,C}
X
current signals V = [V1 , V2 , . . . , VN ]T , I = [I1 , I2 , . . . , IN ]T Inγ = Γ(AIni , AInj ), (10)
are measured from a network with size N the nodal, where i6=j
depending on the number of phases at node i, Vi and Ii can
where, In,max = max{AInA , AInB , AInC } and In,min =
be row vectors of size 1, 2 or 3. In order to characterize the
min{AInA , AInB , AInC }, Γ denotes a thresholding function
waveform properties, we adopt instantaneous properties from:
to measure the difference. Similarly, we can also get Vα , Vβ
sc (t) = s(t) + jH{s(t)} = A(t)ejψ(t) , (5)
and Vγ . If Iβ and Vβ is not zero, there exists unbalance among
where s(t) is the real signal, sc (t) is the complex expression,
the three phases. Then we use Iγ and Vγ to determine the how
A(t) is the instantaneous amplitude (IA) (envelope), ψ(t) is
many phases are affected. In addition, Iα and Vα are used to
the instantaneous phase(IP), H is the Hilbert transform as:
Z∞ measure the absolute changes.
1 s(τ )
H{s(t)} = dτ. (6) C. Identifying the Event Source Region
π t−τ
−∞
Thus, for a three phase current In = [InA , InB , InC ]T , Assuming there are two waveform sets from two nodes, if
where InA = AInA ejψInA (t) . (Vn can be expressed in the it happens, an event may occur in three regions: upstream of
same way.) node n, downstream of node p, and between nodes n and p.
The measurements can be expressed as Mn and Mp . Since we
Algorithm 1 Waveform based Generator Fault Identification and not only have the voltage and current, but also the unbalance
Location Algorithm measurements, Mn can be defined as
1: Input: Waveforms of voltage and current.
Mn = [InA , InB , InC , VnA , VnB , VnC , Iα , Iβ , Iγ , Vα , Vβ , Vγ ]T .
2: Output: Fault properties and location.
3: Event detection based on time series anomaly detection tech- (11)
niques; Then, event location can determined by the comparison
4: Unbalance detection to distinguish single phase, double phase or between ∆Mn and ∆Mp . For example, if the change patterns
three phase fault; of ∆Mn and ∆Mp are the same, the event happens either on
5: Event source region determination based on measurement changes;
the upstream or downstream, but if those are different, the
event happens between nodes n and p.

3
 Fig. 3: Single Phase Short Fault Simulation Results.

could be 8 sensors in the traditional power networks. In our

Fig. 2: IEEE 14 Bus Power Network experiment, we only use 4 sensors from Nodes 1, 6, 7, and 14,
to simulate the limited sensor situation.
V. S IMULATION AND E VALUATION Figure 4 shows the voltages and IAs from the 4 sensors.
Based on the Vγ function, we can know only Phase A has
For evaluation, a 14-bus IEEE standard power network the short circuit event. Although observations from Nodes 7
(Figure 2) is built to investigate the impact of the generator and 14 indicate differences between Phases B and C, the small
faults on electrical waveforms in power networks, which is deviations should be caused be the unbalanced Phase A current
also used to generate a large set of waveform data needed in rather than more than one phase short circuit.
the proposed algorithms. In the simulation model, generators
are simplified as non-ideal three-phase voltage sources, and
the loads are modeled as constant power loads. Meanwhile,
the transmission line reactances are modeled by Simulink
block ’Three-Phase PI Section Line’. And the parameters are
calculated from the IEEE standard 14-bus testbed datasheet.
Specifically, the level of the short fault or the aging degree is
modeled as a short circuit resistance, Rf , as shown in figure 2
(only depict the single phase grounded short fault for example).
Rf is modeled as Rf = 10(5−10∗κaging ) , where κaging denotes
the generator condition aging level.
When two generators highlighted in Figure 2 have aging
stator windings both at 50% level, their voltage and current
signature will change and will likely impact raw electrical
waveforms and signals (e.g.,voltage, current, harmonics, power
factor) in power networks (e.g, voltage, current, etc) due to
waveform and signal propagation. As a result, the voltage and
current waveforms in bus 6 generator and bus from 6 to 11 are
significantly impacted as shown in Figure 3. The traditional Fig. 4: Voltage waveforms and corresponding instantaneous
approaches based on machine sensors, for instance, sensors in amplitudes at Nodes 1, 6, 7 and 14.
bus 6 generator, cannot distinguish whether bus 6 generator
is faulty or not. To address this challenges, we developed an Next, let us locate the root source region. Figure 5 shows
advanced signal processing approach based on strategically the current waveforms measured at Node 6. From Table III,
located waveform senors. The proposed data-driven approach Iβ is the largest for IG6, which means the major unbalance
utilizes waveforms as well as derived signal attributes to source for Node 6 is from the Node 5 direction. (If Node 6
characterize the dynamic relationships among limited sensors to is the fault source, the current unbalance metrics between 6
deduce the actual fault locations, which may not be monitored. and 11, 12, 13 would be similar.) So the fault source could be
Based on the model shown in Figure 2, two generator faults Nodes 1, 2 or 3. However, if Node 2 or 3 is the fault source,
are simulated, which are located at G1 and G8 with 50% aging. the current measured at Node 7 about Bus 4 would display
There are 14 nodes in the grid, and typically all generators more obvious unbalances. While, Node 1 could be the source,
(5) and transformers (3) should have built-in sensors, so there as both its voltage and current are not balanced.

4
 Furthermore, we need to check whether Node 1 is one different scenarios. A hierarchal signal processing technique has
only fault source. If yes, the unbalanced current metrics from been developed, which shows promise in locating fault location
Table III should show largest value from Bus 4 instead of and monitoring the condition, or more specifically identifying
current between Nodes 7 and 8. Thus, for Node 7, the major the aging level. With the reduced number of strategically
fault source is from Node 8, which is also a generator. So, we located waveform sensors, the proposed methodology will
decide Nodes 1 and 8 are two fault sources. Considering the bring the advantages of reducing the cost of the maintenance
space limit, more discussions are omitted. and improving the reliability of the power grid.
R EFERENCES
[1] H. Henao, G. Capolino, M. Fernandez-Cabanas, F. Filippetti, C. Bruzzese,
E. Strangas, R. Pusca, J. Estima, M. Riera-Guasp, and S. Hedayati-Kia,
“Trends in fault diagnosis for electrical machines: A review of diagnostic
techniques,” IEEE Industrial Electronics Magazine, vol. 8, no. 2, pp.
31–42, June 2014.
[2] C. Cecati, Z. Gao, and S. X. Ding, “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,
vol. 62, no. 6, pp. 3757–3767, June 2015.
[3] Z. Gao, C. Cecati, and S. X. Ding, “A survey of fault diagnosis and fault-
tolerant techniques part ii: Fault diagnosis with knowledge-based and
hybrid/active approaches,” IEEE Transactions on Industrial Electronics,
vol. 62, no. 6, pp. 3768–3774, June 2015.
[4] S. Nandi, H. A. Toliyat, and X. Li, “Condition monitoring and fault
diagnosis of electrical motors?a review,” IEEE Transactions on Energy
Conversion, vol. 20, no. 4, pp. 719–729, Dec 2005.
[5] V. Climente-Alarcon, J. A. Antonino-Daviu, A. Haavisto, and A. Arkkio,
“Diagnosis of induction motors under varying speed operation by principal
slot harmonic tracking,” IEEE Transactions on Industry Applications,
vol. 51, no. 5, pp. 3591–3599, Sept 2015.
[6] G. Gentile, S. Meo, and A. Ometto, “Induction motor current signature
Fig. 5: Current waveforms and corresponding instantaneous analysis to diagnostics, of stator short circuits,” in 4th IEEE International
amplitudes at Node 6. Symposium on Diagnostics for Electric Machines, Power Electronics
and Drives, 2003. SDEMPED 2003., Atlanta, GA, USA, Aug 2003, pp.
In addition, based on the current and voltage changes, we 47–51.
estimate the Rˆf 1 = 9.17Ω and Rˆf 8 = 1.35Ω. According to the [7] G. M. Joksimovic and J. Penman, “The detection of inter-turn short
circuits in the stator windings of operating motors,” IEEE Transactions
relation between Rf and κaging in Section V, κG1 aging = 40.3% on Industrial Electronics, vol. 47, no. 5, pp. 1078–1084, Oct 2000.
and κG8
aging = 48.7%. The aging percentage estimation for Node [8] A. Stavrou, H. G. Sedding, and J. Penman, “Current monitoring for
8 is close to the ground truth 50%, while there is still space detecting inter-turn short circuits in induction motors,” IEEE Transactions
on Energy Conversion, vol. 16, no. 1, pp. 32–37, March 2001.
for improvement as κG1 aging is not accurate. [9] W. T. Thomson and M. Fenger, “Current signature analysis to detect
Note that the measurements from Node 14 are not useful in induction motor faults,” IEEE Industry Applications Magazine, vol. 7,
our experiment, so we actually only use 3 nodes to locate the no. 4, pp. 26–34, July 2001.
[10] M. A. S. Nejad and M. Taghipour, “Inter-turn stator winding fault
event source. There are two remarks: (1) limited measurements diagnosis and determination of fault percent in pmsm,” in 2011 IEEE
can be used to monitor the whole network; (2) the observation Applied Power Electronics Colloquium (IAPEC), Johor Bahru, Malaysia,
selection (sensor location) should be considered carefully to April 2011, pp. 128–131.
[11] Y. Liang, “Diagnosis of inter-turn short-circuit stator winding fault in
get the maximum information using minimum sensors and pmsm based on stator current and noise,” in 2014 IEEE International
avoid recording useless information. Conference on Industrial Technology (ICIT), Busan, South Korea, Feb
TABLE III: Unbalance metrics of Nodes 1, 6, 7, 14. 2014, pp. 138–142.
[12] L. L. H. Costa, P. J. A. Serni, and F. P. Maraf?o, “An analysis of
Node # 1 6 7 14 generalized symmetrical components in non sinusoidal three phase
Vβ 0.832 0.328 0.511 0.383 systems,” in XI Brazilian Power Electronics Conference, Praiamar, Brazil,
(1-2) 0.922 (G6) 0.6865 (B4) 0.269 (B14) 0.390 Sept 2011, pp. 502–507.
[13] T. Y. Zheng, S. T. Cha, B. E. Lee, P. A. Crossley, M. Song, and Y. C.
(1-5) 0.825 (6-11) 0.359 (7-8) 0.887 N/A Kang, “Design and evaluation of a protection algorithm for a wind turbine
Iβ
N/A (6-12) 0.181 (7-9) 0.627 N/A generator based on the fault-generated symmetrical components,” in 2011
N/A (6-13) 0.366 N/A N/A 2nd IEEE PES International Conference and Exhibition on Innovative
Smart Grid Technologies, Manchester, UK, Dec 2011, pp. 1–6.
VI. C ONCLUSION [14] S. Altaf, A. Al-Anbuky, and H. GholamHosseini, “Fault diagnosis
in a distributed motor network using artificial neural network,” in
In this paper, an advanced signal processing methodology for 2014 International Symposium on Power Electronics, Electrical Drives,
generator condition monitoring and fault diagnosis has been Automation and Motion, June 2014, pp. 190–197.
developed based on strategically located waveform sensors [15] C. Tsai, W. Lee, E. Nashawati, C. Wu, and H. Lan, “Pmu based generator
parameter identification to improve the system planning and operation,”
in power networks. The impact of the generator short circuit in 2012 IEEE Power and Energy Society General Meeting, San Diego,
fault caused by winding aging and other severe operation CA, USA, July 2012, pp. 1–8.
conditions are discussed, and then analyzed in a great detail. [16] B. Mogharbel, L. Fan, and Z. Miao, “Least squares estimation-based
synchronous generator parameter estimation using pmu data,” in 2015
Meanwhile, an IEEE 14-bus power network is built in the IEEE Power Energy Society General Meeting, Denver, CO, USA, July
MATLAB Simulink to produce a large set of waveform data in 2015, pp. 1–5.