Single User License Agreement

SINGLE USER LICENSE AGREEMENT
THIS IS A LEGALLY BINDING AGREEMENT BETWEEN YOU AND THE ELECTRIC POWER RESEARCH INSTITUTE, INC.
(EPRI). PLEASE READ IT CAREFULLY BEFORE BREAKING OR TEARING THE WARNING LABEL AND OPENING THIS
SEALED PACKAGE.
BY OPENING THIS SEALED PACKAGE YOU ARE AGREEING TO THE TERMS OF THIS AGREEMENT. IF YOU DO NOT AGREE
TO THE TERMS OF THIS AGREEMENT, PROMPTLY RETURN THE UNOPENED PACKAGE TO EPRI AND THE PURCHASE
PRICE WILL BE REFUNDED.
1. GRANT OF LICENSE
EPRI grants you the nonexclusive and nontransferable right during the term of this agreement to use this
report and software only for your own benefit and the benefit of your organization. This means that the
following may use this report and software: (I) your company (at any site owned or operated by your
company); (II) its subsidiaries or other related entities; and (III) a consultant to your company or related
entities, if the consultant has entered into a contract agreeing not to disclose the report and software outside
of its organization or to use the report and software for its own benefit or the benefit of any party other than
your company.
This shrink-wrap license agreement is subordinate to the terms of the Master Utility License Agreement
between most U.S. EPRI member utilities and EPRI. Any EPRI member utility that does not have a Master
Utility License Agreement may get one on request.
2. COPYRIGHT
This report and software, including the information contained therein, are owned by EPRI and are protected
by United States and international copyright laws. You may not, without the prior written permission of EPRI,
reproduce, translate or modify this report and software, in any form, in whole or in part, or prepare any deriv-
ative work based on this report and software.
3. RESTRICTIONS
You may not rent, lease, license, disclose or give this report and software to any person or organization, or
use the information contained in this report and software, for the benefit of any third party or for any pur-
pose other than as specified above unless such use is with the prior written permission of EPRI. You agree
to take all reasonable steps to prevent unauthorized disclosure or use of this report and software. Except as
specified above, this agreement does not grant you any right to patents, copyrights, trade secrets, trade
names, trademarks or any other intellectual property, rights or licenses in respect of this report and software.
4. TERM AND TERMINATION
This license and this agreement are effective until terminated. You may terminate them at any time by
destroying this report and software. EPRI has the right to terminate the license and this agreement immedi-
ately if you fail to comply with any term or condition of this agreement. Upon any termination you may
destroy this report and software, but all obligations of nondisclosure will remain in effect.
5. DISCLAIMER OF WARRANTIES AND LIMITATION OF LIABILITIES
NEITHER EPRI, ANY MEMBER OF EPRI, ANY COSPONSOR, NOR ANY PERSON OR ORGANIZATION ACT-
ING ON BEHALF OF ANY OF THEM:
(A) MAKES ANY WARRANTY OR REPRESENTATION WHATSOEVER, EXPRESS OR IMPLIED, (I) WITH
RESPECT TO THE USE OF ANY INFORMATION, APPARATUS, METHOD, PROCESS OR SIMILAR ITEM DIS-
CLOSED IN THIS REPORT AND SOFTWARE, INCLUDING MERCHANTABILITY AND FITNESS FOR A PAR-
TICULAR PURPOSE, OR (II) THAT SUCH USE DOES NOT INFRINGE ON OR INTERFERE WITH PRIVATELY
OWNED RIGHTS, INCLUDING ANY PARTY’S INTELLECTUAL PROPERTY, OR (III) THAT THIS REPORT AND
SOFTWARE IS SUITABLE TO ANY PARTICULAR USER’S CIRCUMSTANCE; OR
(B) ASSUMES RESPONSIBILITY FOR ANY DAMAGES OR OTHER LIABILITY WHATSOEVER (INCLUDING
ANY CONSEQUENTIAL DAMAGES, EVEN IF EPRI OR ANY EPRI REPRESENTATIVE HAS BEEN ADVISED OF
THE POSSIBILITY OF SUCH DAMAGES) RESULTING FROM YOUR SELECTION OR USE OF THIS REPORT
AND SOFTWARE OR ANY INFORMATION, APPARATUS, METHOD, PROCESS OR SIMILAR ITEM DIS-
CLOSED IN THIS REPORT AND SOFTWARE.
6. EXPORT
The laws and regulations of the United States restrict the export and re-export of any portion of this report
and software, and you agree not to export or re-export this report and software or any related technical data
in any form without the appropriate United States and foreign government approvals.
7. CHOICE OF LAW
This agreement will be governed by the laws of the State of California as applied to transactions taking place
entirely in California between California residents.
8. INTEGRATION
You have read and understand this agreement, and acknowledge that it is the final, complete and exclusive
agreement between you and EPRI concerning its subject matter, superseding any prior related understand-
ing or agreement. No waiver, variation or different terms of this agreement will be enforceable against EPRI
unless EPRI gives its prior written consent, signed by an officer of EPRI.
9807070
9807070
On-line Detection of Shorts in
Generator Field Windings
TR-114016
Final Report, November 1999
EPRI Project Manager

J. Stein
EPRI • 3412 Hillview Avenue, Palo Alto, California 94304 • PO Box 10412, Palo Alto, California 94303 • USA
800.313.3774 • 650.855.2121 • askepri@epri.com • www.epri.com
9807070
DISCLAIMER OF WARRANTIES AND LIMITATION OF LIABILITIES
THIS DOCUMENT WAS PREPARED BY THE ORGANIZATION(S) NAMED BELOW AS AN
ACCOUNT OF WORK SPONSORED OR COSPONSORED BY THE ELECTRIC POWER RESEARCH
INSTITUTE, INC. (EPRI). NEITHER EPRI, ANY MEMBER OF EPRI, ANY COSPONSOR, THE
ORGANIZATION(S) BELOW, NOR ANY PERSON ACTING ON BEHALF OF ANY OF THEM:
(A) MAKES ANY WARRANTY OR REPRESENTATION WHATSOEVER, EXPRESS OR IMPLIED, (I)

WITH RESPECT TO THE USE OF ANY INFORMATION, APPARATUS, METHOD, PROCESS, OR
SIMILAR ITEM DISCLOSED IN THIS DOCUMENT, INCLUDING MERCHANTABILITY AND FITNESS
FOR A PARTICULAR PURPOSE, OR (II) THAT SUCH USE DOES NOT INFRINGE ON OR
INTERFERE WITH PRIVATELY OWNED RIGHTS, INCLUDING ANY PARTY'S INTELLECTUAL
PROPERTY, OR (III) THAT THIS DOCUMENT IS SUITABLE TO ANY PARTICULAR USER'S
CIRCUMSTANCE; OR
(B) ASSUMES RESPONSIBILITY FOR ANY DAMAGES OR OTHER LIABILITY WHATSOEVER

(INCLUDING ANY CONSEQUENTIAL DAMAGES, EVEN IF EPRI OR ANY EPRI REPRESENTATIVE
HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES) RESULTING FROM YOUR
SELECTION OR USE OF THIS DOCUMENT OR ANY INFORMATION, APPARATUS, METHOD,
PROCESS, OR SIMILAR ITEM DISCLOSED IN THIS DOCUMENT.
ORGANIZATION(S) THAT PREPARED THIS DOCUMENT
Computational Intelligence Applications Laboratory
ORDERING INFORMATION
Requests for copies of this report should be directed to the EPRI Distribution Center, 207 Coggins
Drive, P.O. Box 23205, Pleasant Hill, CA 94523, (800) 313-3774.
Electric Power Research Institute and EPRI are registered service marks of the Electric Power
Research Institute, Inc. EPRI. POWERING PROGRESS is a service mark of the Electric Power
Research Institute, Inc.
Copyright © 1999 Electric Power Research Institute, Inc. All rights reserved.
9807070
CITATIONS
This report was prepared by
Computational Intelligence Applications Laboratory

University of Washington
Seattle, WA
Principal Investigators
M. El-Sharkawi
R. J. Marks
Project Advisor
I. Kerszenbaum, Edison Mission Energy
This report describes research sponsored by EPRI and the National Science Foundation.
The report is a corporate document that should be cited in the literature in the following manner:
On-line Detection of Shorts in Generator Field Windings, EPRI, Palo Alto, CA, and the National
Science Foundation: 1999. TR-114016.
iii
9807070
9807070
REPORT SUMMARY
Shorted turns in the field winding of large turbogenerators are a common problem; their
detection and localization have remained elusive. There are many techniques for detection of
shorted turns; but all of these methods are expensive to implement; and none can localize the
shorted turn. The main objective of the work in this project is to develop an intelligent system
with associated hardware to implement the traveling wave technique, an approach that can both
detect and localize shorted winding on-line without the installation of a sensing coil.
Background
Shorts in turbogenerators occur primarily from incessant pounding of the rotor copper
conductors while the machine is turning in low gear. This low speed operation is designed to
avoid the deformation of the shaft that occurs when the rotor remains stationary in the bore for
long periods of time. The pounding of the copper conductors results in the accumulation of
copper powder within slots. When the machine is subsequently energized, the copper dust causes
arcing between the turns in the slot. Over time, a full short circuit between turns may result.
Broken rotor conductors and water intrusion may also cause short turns. In many instances, the
rotor short turns are speed dependent so that the fault tends to disappear once the machine is
brought to standstill. This makes the determination of such a fault difficult. Many techniques
have been reported for detection of shorted turns, most of which require some modification of
the stator windings or placement of flux coils in the air gap. One method is based on sensing the
increased mechanical vibration. A second method is based on installing pick-up coils inside the
machine to detect the shorts by monitoring the induced voltage on the coil. Yet another method
is based on reconfiguring the structure of the windings in such a way that the short can be detect
by flux asymmetry. None of these methods can localize a short.
Objective
To develop and test an intelligent system with associated hardware to implement the traveling
wave technique for detection and localization of shorted windings in turbogenerators.
Approach
The project team designed, built, and tested a system to detect and localize shorted windings in
turbogenerators. The team built hardware including a signal generator and signal detector to send
two identical travelling waves into both ends of a turbogenerator simultaneously and capture the
resulting signature signals. The team developed a statistical technique, the novelty filter, to
identify anomalous signature signals indicative of shorted windings. They built and trained a
fuzzified neural network to localize the detected shorts when the turbogenerator is running. The
team conducted laboratory tests to verify the ability of the system to identify and localize shorts.
v
9807070
Results
In this detection and localization system, travelling waves are sent simultaneously through the
windings from opposite ends of the coil. The difference between the reflected signals is
observed. The signals should be identical unless a short is present. In order to automate the
analysis of the signals, the system incorporates a novelty filter that identifies outliers in the data
that may indicate a short. Multi-layer perceptron neural network technology and efficient
training algorithms are used that can localize detected shorts. Not only is this method effective,
but the equipment needed to apply the test is external to the coil. A sensing coil does not need to
be installed, and the test can be performed while the turbogenerator is on-line.
The detection and localization methods developed in this project were lab tested using the
prototype hardware. For the detection tests, a set of signatures was gathered at several different
operating speeds and excitation currents. To observe the effects of shorted windings, 27 shorts
were made across adjacent windings covering all slots, poles, and coils. Several of these shorts
were near the center and the edges of the coil. All faults were successfully detected. For the
localization tests, a total of 450 signatures were collected for several shorts—70% of them are
used for training and the rest for testing. A neural net with two hidden layers, seven nodes each,
five input neurons, and one output neuron was used. After training, the neural net was tested for
135 shorted data. Training and testing results were remarkably good.
EPRI Perspective
To further refine the system, future works in this area should include detection of intermittent
and multiple shorts, establishment of a signature signals database, and a large-scale field test to
verify the effectiveness of the developed system. The short detection should be tested for
operational and excited machines. For localization, the test should be done when the machine is
out of service for maintenance.
TR-114016
Keywords
Turbogenerators
Neural nets
Artificial intelligence
Fault detection
vi
9807070
EPRI Licensed Material
CONTENTS
1 PROBLEM FORMATION AND SYSTEM FUNCTIONAL SPECIFICATION ........................ 1-1

1.1 Problem Formulation .................................................................................................... 1-1
1.2 General Operational Specifications............................................................................... 1-2
2 FULLY FUNCTIONAL SYSTEM DESIGN ........................................................................... 2-1

2.1 Basic Blocks ................................................................................................................. 2-1
2.2 Test system .................................................................................................................. 2-1
2.3 Hardware Design.......................................................................................................... 2-4
2.3.1. Signal Generator.................................................................................................. 2-5
2.3.2. Signal Detector .................................................................................................... 2-7
2.3.3 Problems and Modifications .................................................................................. 2-9
2.4 Hardware Design and Requirements for Field test........................................................ 2-9
3 SIGNATURE SIGNALS AND OUTLIER DETECTION......................................................... 3-1

3.1 Twin and In-Out Signature Signals ............................................................................... 3-1
3.2 Outlier Detection........................................................................................................... 3-2
4 NOVELTY FILTER FOR SHORT DETECTION ................................................................... 4-1

4.1 Computation of Detection Threshold............................................................................. 4-3
4.2 Mahalanobis Distance .................................................................................................. 4-6
4.3 Lab Evaluation.............................................................................................................. 4-7
5 NEURAL NETWORK FOR SHORT LOCALIZATION.......................................................... 5-1

5.1 Fuzzified Neural Network.............................................................................................. 5-1
5.2 Considerations for Operational Generators................................................................... 5-5
5.2.1 Brush noise characterization ................................................................................. 5-6
5.2.2 Field Verification ................................................................................................... 5-8
6 LABORATORY TEST.......................................................................................................... 6-1
vii
9807070
6.1 Short Detection............................................................................................................. 6-3

6.2 Short Localization ......................................................................................................... 6-8
7 CONCLUSIONS AND FUTURE WORK .............................................................................. 7-1
8 BILIOGRAPHY .................................................................................................................... 8-1
viii
9807070
LIST OF FIGURES
Figure 2-1 Basic Setup............................................................................................................ 2-2

Figure 2-2 Outline of Procedure .............................................................................................. 2-2
Figure 2-3 Photo of the prototype rotor.................................................................................... 2-2
Figure 2-4 Winding configuration of test rotor.......................................................................... 2-3
Figure 2-5 Drive system for the setup ..................................................................................... 2-3
Figure 2-6 Excitation slip rings of the test rotor........................................................................ 2-4
Figure 2-7 Block Diagram and Photo of Hardware Setup ........................................................ 2-4
Figure 2-8 Main Circuit diagram of signal generator ................................................................ 2-6
Figure 2-9 Waveforms at various points in the signal generator circuit .................................... 2-6
Figure 2-10 Power supply circuit for signal generator.............................................................. 2-7
Figure 2-11 Block diagram of signal detector circuit ................................................................ 2-8
Figure 2-12 Signal impressed on device under test................................................................. 2-8
Figure 2-13 Typical signature signals for stopped test rotor .................................................... 2-8
Figure 3-1 Twin Signal Sensing............................................................................................... 3-1
Figure 3-2 Impedance Match in Conjunction with In-out Signal Sensing ................................. 3-2
Figure 4-1 Simplified view of the projection of a signature onto the hyperellipse. The
measured signature is shown to be projected onto the hyperellipse but to be
significantly removed from the projection plane ............................................................... 4-5
Figure 4-2 Typical signature signals for different rotor speeds under no load .......................... 4-8
Figure 4-3 Signature signals of healthy rotor obtained from a rotor operating at full speed
(top) and stopped (bottom) under different excitations levels........................................... 4-9
Figure 4-4 Signature signals for a stopped rotor with shorted turns....................................... 4-10
Figure 4-5 False alarm, and miss rate for different thresholds.............................................. 4-11
Figure 5-1 Short turn localization procedure............................................................................ 5-1
Figure 5-2 Membership function for the fuzzy mappings ......................................................... 5-2
Figure 5-3 Samples of signature signals ................................................................................. 5-3
Figure 5-4 Features extracted from signature signal ............................................................... 5-4
Figure 5-5 Brush noise measurement showing noise spikes and power supply ripple............ 5-7
Figure 5-6 Brush noise measurement showing the periodic nature of the noise spikes ........... 5-7
Figure 5-7 Power spectrum of brush noise where each measurement was the average of
100 oscilloscope sweeps................................................................................................. 5-8
Figure 5-8 Power spectrum of brush noise.............................................................................. 5-8
ix
9807070
Figure 5-9 Plot of signature signal taken directly from the slip rings (solid line) and the
signature signal taken through the brushes (dotted line).................................................. 5-9
Figure 5-10 Signature signals for various rotation rates ........................................................ 5-10
Figure 5-11 Signature signals for two different machines. The measurement leads were
connected directly to the slip rings................................................................................ 5-10
Figure 6-1 Typical Signatures for Different Speeds (Twin Signal)............................................ 6-1
Figure 6-2 Typical Signatures for Different Speeds (In-Out Signal) ......................................... 6-2
Figure 6-3 Shorted Turn Signatures for Stopped Rotor (Twin Signal)...................................... 6-2
Figure 6-4 Shorted Turn Signatures for Stopped Rotor (In-out Signal) .................................... 6-3
Figure 6-5 ROC Curves for the Running Rotor without DC-Excitation (Twin Signal)............... 6-4
Figure 6-6 ROC Curves for the Running Rotor without DC-Excitation (In-out Signal) .............. 6-4
Figure 6-7 ROC Curves for DC-Excitation (Twin Signal) ......................................................... 6-5
Figure 6-8 ROC Curves for the DC-Excitation (in-out Signal) .................................................. 6-5
Figure 6-9 The Elliptic Boundary Technique using Twin-Signal method .................................. 6-6
Figure 6-10 The Elliptic Boundary Technique using in-out method.......................................... 6-7
Figure 6-11 The Elliptic Boundary Technique using Twin-Signal method ................................ 6-7
Figure 6-12 The Elliptic Boundary Technique using in-out method.......................................... 6-8
Figure 6-13 The Training Result.............................................................................................. 6-9
Figure 6-14 The Testing Result ............................................................................................. 6-10
x
9807070
LIST OF TABLES
Table 4-1 Detection results for voltage-excited running rotor................................................ 4-10

Table 5-1 Field test results ..................................................................................................... 5-5
Table 6-1 Detection Results for Running Rotor Using Twin Signal and In-out Signal ............. 6-3
Table 6-2 Localization Results for Running Rotor Using Neural Network ............................... 6-9
xi
9807070
9807070
1
PROBLEM FORMATION AND SYSTEM FUNCTIONAL
SPECIFICATION
1.1 Problem Formulation
Shorted turns in the field winding of large turbogenerators are a common problem whose
detection and localization have remained elusive. Shorts occur primarily from incessant
pounding of the rotor copper conductors while the machine are turning in low gear. This low
speed operation is designed to avoid the deformation of the shaft that occurs when the rotor
remains stationary in the bore for long periods of time.
The pounding of the copper conductors results in the accumulation of copper powder within
slots. When the machine is subsequently energized, the copper dust causes arcing between the
turns in the slot. Over time, a full short circuit between turns may result. Broken rotor
conductors and water intrusion may also cause short turns.
In many instances, the rotor short turns are speed dependent, i.e., the fault tends to disappear
once the machine is brought to standstill. This makes the determination of such a fault difficult
[1-9].
Many techniques have been reported for detection of shorted turns, most of which require some
modification of the stator windings or placement of flux coils in the air gap. One method is
based on sensing the increased mechanical vibration [4]. A second method is based on installing
pick-up coils inside the machine to detect the shorts by monitoring the induced voltage on the
coil [1]. Yet another method is based on reconfiguring the structure of the windings [2] in such a
way that the short can be detect by flux asymmetry. All of these methods are expensive to
implement, and none can localize the shorted turn [3].
An effective method for shorted turn detection and localization is based on the traveling waves
[5-15]. Two traveling waves are sent through the windings from opposite ends of the coil. The
difference between the reflected signals, dubbed the twin signals, is observed. The twin signals
should be identical unless a short is present. An expert can analyze the signals and recognize the
existence of short. Not only is this method effective, the equipment needed to apply the test is
external to the coil. No time-consuming installation of the equipment or the special design of the
winding in the turbogenerator is needed.
The main objective of the work in this project is to develop an intelligent system with associated
hardware to implement the traveling wave technique for shorted winding detection and
localization. The technique does not require an insertion of probes or special coils inside the
machine, and can be used on-line.
1-1
9807070
Problem Formation and System Functional Specification
Among the main challenges of this project are acquisition and interpretation of traveling wave
data in the presence of shorted windings. No data is available for industrial rotors with shorted
turns. It is also obviously impractical to generate such data from a commercial turbine-generator.
An alternative, model-based generation of data, would be prohibitively complex and difficult to
verify.
A viable and attractive alternative methodology for shorted windings detection not requiring data
or a model of the shorted windings is the novelty detector. The novelty detector wraps a closed
decision surface around a given set of data from a “healthy” rotor. All data points within the
surface are recognized as “healthy”, while data points outside are related to something new (i.e
novel) or abnormal in the rotor, and are indicators of a shorted turn. The novelty detector is
remarkably effective for accurately detecting shorted turns [9].
The on-line short localization proves to be a more difficult problem than the off-line short
detection. The main difficulty is specification of a desirable localization metric to indicate the
short location. Even when specified, a technique is needed to relate the original rotor signals with
the localization metric. In this research, several localization metrics were studied, inxluding the
resistance between the short location and one end of the rotor winding. The localization metrics
are used to train a neural network using the backpropagation training method. In order to avoid
the curse of dimensionality and improve the effectiveness of the neural network, feature
extraction is used to reduce the number of inputs of the neural network [15-18].
1.2 General Operational Specifications
There are several design considerations in this project:
1. Injected signal frequency. The developed method is based on injecting two identical high
frequency signals at both ends of the rotor field windings. The frequency of the injected
1
signals is selected at a rate no greater than , where τ which is the traveling time of the
10 τ
field winding. The interference between the falling edge of the injected signal and the
reflected wave is then essentially eliminated.
2. On line noise. The detection of faults on-line is much more difficult due to the presence of
noise caused by
• Brushes,
• Mechanical variations during rotation, and
These unwanted contributions are effectively filtered by separating the injected signals
from the noise using frequency domain filtering.
3. Asymmetric air gaps. The air gap may not be symmetrical, especially in older machines.
This may cause a low frequency stochastic component to appear in the signature signals.
Removal of the rotational effects is complex, but can be accomplished using the novelty
filter.
1-2
9807070
Problem Formation and System Functional Specification
4. False conduction paths. For on-line applications, the excitation system provides an second
unwanted path for the injected signals. This added path can change the characteristics of the
reflected signals. This effect is successfully addressed by developing the detection system
using measured data wherein these extraneous signal components are present. The resulting
novelty detector then adapts to the distorted signal.
1-3
9807070
9807070
2
FULLY FUNCTIONAL SYSTEM DESIGN
2.1 Basic Blocks
The circuitry for generating twin signals is depicted in Figure 2-1. The basic functional blocks
are shown in Figure 2-2. A signal generator is used to send two signals to the symmetrical
circuit. The two identical signals are injected into the winding from both sides. The reflected
waveforms are received, subtracted and processed. The subtracted signal serves as the fingerprint
of performance and is dubbed the signature signal. The twin signals are acquired by a PC-based
computer through a general purpose interface board (GPIB). This high frequency sampling
device is used to ensure that the entire signature signal is captured. The shape of the signature
signal is used to perform two functions: 1) detect the existence of a shorted turn; and 2) localize
the short. The frequency of the injected signal must be selected to ensure no interference
between the falling edge of the injected signal and the reflected wave.
2.2 Test system
To test the proposed methodology, a rotor was constructed in the Computational Intelligence
Applications (CIA) Lab at the University of Washington. The rotor has 4 poles and 12 slots with
two concentric coils per pole. A photo of the rotor is shown in Figure 2-3 and the winding
configuration is shown in Figure 2-4. The inner concentric coil has 36 turns and the outer coil
has 18 turns. Thus, each slot has 36 conductors. Each of the 4 windings are connected in series
with appropriate polarity. To enable creation of shorts, one end of the coils is left accessible. The
insulation on several adjacent turns can be removed to allow for creation of shorts externally. Of
the 36 turns in the inner coil, 8 groups of alternate turns can be shorted and 4 groups of the 18
outer turns can also be shorted. This permits sufficient variation in the location of the fault. A
DC motor connected using a belt-pulley drives the rotor at any desired speed. A photo of the
drive system is shown in Figure 2-5. Excitation in the form of a DC or AC voltage can be applied
to the rotor through two slip rings connected to the windings as shown in the photo of Figure 2-6.
Thus, the test rotor can simulate the effect of rotation of an excited winding. Since the windings
are required to be available for inducing artificial shorts, a stator was not built. Consequently, the
effect of stator flux on the rotor windings and the signature signal is not available.
2-1
9807070
Fully Functional System Design
Winding
Symmetrical
Signal Generator
Circuit
GPIB
Figure 2-1
Basic Setup
Sending
Block Receiving
Block
Signal
Generator
Receiving
Sending Block
Block
Signal Difference
GPIB
Oscilloscope
Emulator
Fuzzified NN
Figure 2-2
Outline of Procedure
Figure 2-3
Photo of the prototype rotor
2-2
9807070
Pole 1
Inner Windings
Pole 4
Pole 2
Slots
Pole 3 Outer Windings
Figure 2-4
Winding configuration of test rotor
Figure 2-5
Drive system for the setup
2-3
9807070
Figure 2-6
Excitation slip rings of the test rotor
2.3 Hardware Design
As shown in Figure 2-7, the circuit in Figure 2-1 can be modified for on-line applications when
the rotor is excited externally. The winding under test is shown as excited by a separate DC
source. The signal generator applies twin signal pulses to the device under test through two series
capacitors that serve to isolate the high winding voltage from the detection circuits. The circuit is
thus designed for use while the winding under test is in normal operation. The reflected signals
obtained at the winding terminals are fed to the signal detector through the series capacitors. The
signal detector filters these signals, finds the difference between the two and amplifies the
difference for further signal processing.
winding under test
Signal Signal
Gen. detector
120 V AC
Sharp edge signals
Figure 2-7
Block Diagram and Photo of Hardware Setup
2-4
9807070
The output of the signal detector circuit is called the signature signal and reflects the current
state of the winding. The signature signal is captured using either a digital storage oscilloscope, a
custom digital scope PC add-on card, or GPIB. It is then compared to the a template signature
obtained from the same winding in the past. If there is a significant deviation between the
template and the signature signal, a novelty is declared, signifying the existence of a short
somewhere in the winding.
The signal generator circuit produces two low voltage pulses which are applied to both ends of
the winding. To ensure minimum interference with the main circuit and to increase detection
sensitivity with respect to pulse magnitude, the falling edge of the pulses needs to be very sharp.
These pulses should be applied at or near the zero crossing of the winding voltage if the winding
is excited by an AC source. Thus, the design specifications for the signal generator circuit are:
• generate a small duration periodic pulse with a sharp falling (or rising) edge,
• duplicate the first edge of this pulse exactly for application to both ends of the winding, and
• generate the pulses at or near the zero crossing of the AC voltage for an AC excited winding.
This requires synchronization of the applied periodic pulse with the system AC voltage.
As shown in Figure 2-7, the signal generator circuit has a 120 V AC input and generates identical
fast falling edge pulses at its outputs. These pulses are applied to the winding under test through
series capacitors that also filter the effect of the winding resistance on the signal generator
circuit. This is in addition to their function of isolating the DC voltage from the signal generator
and detector circuits. For maximum sensitivity, the signals are generated at every zero crossing
of the AC waveform, that is, once every half cycle. To ensure uniformity of the circuit for all
applications and for ease of operation, this synchronization with the AC source voltage is also
used for machines with DC windings.
The signal detector circuit amplifies the small difference in the reflected waveforms. Since the
technique looks for changes in the signature signals, it is imperative that the amplification be
highly repetitive and free from drift so as to minimize sensing errors. The detector circuit
therefore utilizes:
• a high pass filter to remove the DC offset or the AC voltage on which the twin signals are
superimposed and
• a low offset, drift-free differential amplifier to accurately amplify these filtered signals.
2.3.1. Signal Generator

Figure 2-8 shows a detailed block diagram of the signal generator circuit. Given the design
specifications of the circuit as outlined in the previous section, the zero crossings of the input AC
voltage waveform need to be sensed so as to generate the short negative pulses. The input
voltage is filtered using a zero phase shift analog low pass filter to get rid of any high frequency
noise present on the supply voltage. This avoids spurious zero crossings from being sensed. The
waveforms at point’s b, c and d are shown in Figure 2-9. The output at point b is a sinusoid with
the same phase angle as the waveform at point a. Two comparators form a zero-crossing detector
(ZCD) for this filtered sine waveform. The output of one comparator goes low at the positive
zero crossing of the sine wave while that of the output of the other comparator goes low at the
negative zero crossing.
2-5
9807070
b
T1
low c
120 V a pass ZCD
AC input COMP
filter
Vref
+Vcc
d
555 Astable
multivibrator
e
-Vcc
Vva
Twin pulse
output
Figure 2-8
Main Circuit diagram of signal generator
b
Voltage
time
c
Voltage
time
d
Voltage
time
e
Voltage
time
Figure 2-9
Waveforms at various points in the signal generator circuit
A simple RC derivative circuit generates a sharp falling edge at each comparator output. These
two edges are added by two diodes to obtain a falling edge at every zero crossing at point c. This
voltage is compared with a small DC voltage, Vref using another comparator COMP to obtain
short positive pulses at every zero crossing at point d. These short pulses are used as enabling
signals for an astable multivibrator, which uses a 555 timer. The positive pulses at d are made
larger than the positive pulse of the multivibrator but smaller than its period so that only one
positive pulse with a sharp rising edge is produced. This output drives a transistor totem-pole
producing a bipolar output at e with a small duty cycle, which, in turn, drives the main switching
2-6
9807070
MOSFET to create a very sharp falling edge at each zero crossing. The MOSFET drain is
connected through resistors to the winding under test to obtain identical waveforms. Although,
the electronic components introduce a small delay between the zero crossings and the falling
edge of the output pulses, it is not a detriment since the need is for a constant repetitive delay
throughout the operation of the circuit. Low-offset operational amplifiers and low drift discrete
components are used to ensure repeatability of the waveforms.
Other blocks in the circuit are shown in Figure 2-10 and include
• a linear power supply for the main circuit components of Figure 2.8, and
• a multi-output isolated buck converter.
T2
+V cc
120 V power
a.c. input supply -V cc
Vvar
555 Astable Tap
multivibrator changing
switch
Figure 2-10
Power supply circuit for signal generator
The multiple output buck converter generates serves as the main power supply for the switching
MOSFET of Figure 2-8. It is implemented using an astable multivibrator, a MOSFET and a
transformer with taps for various turn ratios. The multivibrator generates the gate drive for the
MOSFET with a switching frequency of about 40 kHz. A selector switch sets the desired tap on
the transformer to generate the variable voltage. This variable voltage source, Vvar, is used as the
supply for the main switching MOSFET of Figure 2-8. A variable voltage source is required for
changing the amplitude of the twin pulses produced so that a proper magnitude can be selected
depending on the winding being tested. Since this variable voltage supply is isolated from the
input AC source, the twin signals can be applied to any winding to which additional DC or AC
voltage has been applied. The twin signals float on the voltage applied to the winding under test.
2.3.2. Signal Detector
Figure 2-9 shows an internal block diagram of the signal detector circuit. The transformer serves
to isolate the input voltage from the circuit, so that the detector circuit can also float on any DC
or AC voltage. A pair of diodes, connected in anti-parallel, clamps the output of the transformer
to a low value. This bipolar clamped voltage is fed to a low offset, low drift, high bandwidth
operational amplifier configured as a differential amplifier. A potentiometer is used to remove
the offset of the amplifier and another is used to set the desired gain close to a value of 10. The
transformer leakage inductance coupled with the capacitance of the clamping diodes filters some
of the high frequencies in the signal. However this filtering is negligible and mainly serves to
reduce the noise in the reflected signals.
2-7
9807070
reflected external
signals
Differential ±7.5V
amplifier supply
clamping
diodes
Figure 2-11
Block diagram of signal detector circuit
Figures 2-12 and 2-13 show actual output waveforms for the two circuits captured using a digital
oscilloscope. Figure 2-12 shows the pulse generated by the signal generator circuit (top) with a
zoomed version of the signal (bottom). Figure 2-13 shows the amplified difference output from
the detector circuit.
15
Signal (V)
10
0
0 2 4 6 8 10
Time (ms)
20
Signal (V)
10
-10
0 20 40 60 80 100
Time (microseconds)
Figure 2-12
Signal impressed on device under test
1
Without short
0.8 With short
0.6
0.4
0.2
Signal (V)
-0.2
-0.4
-0.6
-0.8
-1
0 2 4 6 8 10 12 14 16 18 20
Time ( µseconds)
Figure 2-13
Typical signature signals for stopped test rotor
2-8
9807070
2.3.3 Problems and Modifications
The amplified difference between the reflected twin signals is captured using either a digital
storage oscilloscope or a custom digital scope PC add-on card. The amplified signal is triggered
at the instant at which the twin signals are applied to the rotor. The applied signal is fed to an
opto-coupler and the isolated signal is used for triggering. However, the delay introduced by the
opto-coupler varies as a function of environmental factors by approximately 800 ns. Since, the
signal range of interest has a duration of only a few microseconds, the variation in the triggering
delay is unacceptable. To overcome this problem, the signal at the MOSFET drain (Figure 2-8) is
directly used as a falling edge triggering signal for the oscilloscope. This requires that either the
machine winding under test should be isolated from the supply neutral or that the oscilloscope
ground should not be connected to the neutral. This is not restrictive since the PC add-on scope
card is isolated from the supply neutral.
Care is taken to ensure that there is no ground loop and that each part is grounded with only one
wire. This ensures that there is no circulating ground current due to variation in potential. Cables
are of minimum necessary length and are shielded wherever possible to minimize the induced
emf generated due to the high frequency twin signals. The multi-output isolated forward
converter causes some differential noise to be added to the twin signals with respect to the circuit
ground. This signal gets amplified due to minor asymmetries and the differential amplifier in the
signal detector. This problem can be solved by careful compensation of the stray inductance in
the high frequency path of the converter.
The signal detector amplifier must be carefully tuned to achieve the high bandwidth with
negligible drift over time. The novelty detection technique is based on data collected from
healthy rotor signatures only. It detects shorts whenever there is a significant change in the
signature signal. All components in the hardware setup are hence required to maintain the same
waveform throughout a large time period. This ensures that any variation in the signature due to
changes in the impressed twin signals does not create a false detection alarm.
2.4 Hardware Design and Requirements for Field test
The hardware is designed to operate in the field while the generator is excited and running. A
separation is designed to isolate the low voltage injected signals from the high voltage excitation
system. The circuit is designed to operate for a DC excitation system, or rectified exciters. The
only requirement of the circuit is access to the slip rings of the rotor windings. Since the test
signals are very low in voltage, they do not interfere with the normal operation of the exciters.
The circuit is small in size and is built in a 3’ NEMA enclosure. Additional equipment such as a
digital scope is needed during the installation process of the system.
The circuit is designed to operate continuously if needed, and is fail safe. The system can be
monitored by a local system or remotely. The remote control feature requires the availability of
a communication link at the site.
2-9
9807070
9807070
3
SIGNATURE SIGNALS AND OUTLIER DETECTION
With the twin signal method, the signature signal is the difference between the reflected signals
at both sides of the windings. Although very effective, the twin signal approach is not applicable
when shorts are close to the center of the rotor windings. In such a case, for symmetric
windings, the reflected twin signals are ideally identical and the signature signal, equal to the
difference of the twin signals, is zero. The same signature signal results when there are no
shorted windings. To overcome this weakness in the twin signal sensing method, an in-out
signal sensing method has been developed.
3.1 Twin and In-Out Signature Signals
The signature signal using the twin signal method is shown in Figure 3-1. The signature signal is
the difference between the two reflected signals [1-5]. When shorted turns occur near the point
of symmetry, the signature signal is difficult to detect.
Sending signal Receiving signal
Sending signal
B
A-B
Figure 3-1
Twin Signal Sensing
To avoid this problem, a method based on impedance matching is proposed. The technique is
described in Figure 3-2. Here, an RC impedance match is connected to one end of the rotor
winding. By adjusting the resistance and/or capacitance of the RC circuit, we can reduce or
eliminate the reflected wave at a specified frequency. When the two signals (A&B) are applied
as shown in the figure, the signature signal is (A-C). Where C is the signal at the matching
impedance terminal. (A-C) is also referred to as in-out signal. By tuning the capacitance and/or
resistance, the energy of the signature can be increased.
3-1
9807070
Signature Signals and Outlier Detection
If the matching impedance is selected for a healthy rotor, any short in the rotor windings creates
a higher energy detectable signature signal. This is true even if the short is in the middle of the
windings.
Sending signal A
Sending signal B
A-C
Figure 3-2
Impedance Match in Conjunction with In-out Signal Sensing
3.2 Outlier Detection
To eliminate significant interference in the measurements and corresponding detection

performance, an outlier detection algorithm is developed to dismiss freakish measurements in the
establishment of the novelty boundary. The interference causing these outliers is due primarily
to problems with data acquisition equipment, grounding, and irregular electromagnetic
transients. When an average is taken for a number of signature signals, the interference biases
the result.
To solve this problem, an outlier detection algorithm is used to filter signals deviating
significantly from a set of "typical" signals. First, let us assume that x n is the signature

signal, n = 1 N . For this set, the mean is calculated by
N
x = N1 ∑ xn (3.1)
n=1
The standard deviation is estimate by
N
s 2 = N1−1 ∑ (x n − x )T (x n − x ) (3.2)
n =1
3-2
9807070
Signature Signals and Outlier Detection
Every signal that satisfies the condition in Equation (3.3) is considered an outlier and is
discarded as a viable signal.
(xi − x )T (xi − x ) > 3s (3.3)
If the signal is not an outlier, the mean is recalculated, and the next signal is checked.
3-3
9807070
9807070
4
NOVELTY FILTER FOR SHORT DETECTION
A novelty filter, as defined by Kohonen [10], is a system that extracts the new anomalous or
unfamiliar part of the input data. With novelty detection, the status quo signal “fingerprint” of
healthy operation is monitored. If the fingerprint samples remain in the region normally
prescribed for healthy operation, there is no reason for alarm. If the fingerprint drifts
significantly from the status quo, however, something novel has occurred. Degradation, i.e. the
forming of a winding short, is then suspected.
Novelty detection is useful in cases where data corresponding to “unhealthy” system conditions
is unavailable. Such is the case for detection of shorted windings. Obtaining signature signals
for operating rotors with shorted windings is prohibitively expensive.
Novelty detection can be conceptually viewed as a method of grouping all signals of a given
training set and comparing future samples with this group. The underlying assumption is that the
training set is statistically representative of all status quo operating conditions and that the
system from which the samples are obtained is not time variant. View each healthy signal as a
point is signal space. A surface is imposed around these points. If a new signal lies within this
surface, it is status quo. Otherwise, it is novel.
Surfaces that can be placed around the healthy data include

• a spherical boundary,
• an elliptical boundary,
• a rectangular boundary formed by the extrema of the data, or min-max surface, and
• nearest neighbor boundaries.
In the first two methods, incoming signatures are compared to a given prototype signature. The
result is declared novel if the signal exceeds a given distance from this prototype. For the
spherical boundary, the distance is a standard Euclidean measure and is therefore equivalent to a
matched filter. When an elliptical boundary is determined along the data’s eigenvectors, the
Mahalanobis distance from the prototype is being used. (The Euclidean distance is a special case
corresponding to equal eigenvalues.)
For the min-max technique, the smallest possible box containing all of the healthy data is used.
The dimensions of the box are determined by the minima and the maxima of the signature signal
elements. This is equivalent to finding upper and lower bounds for all of the healthy data. These
two signals uniquely define the box by its vertices. After the box is defined, each linear
dimension of the box may be proportionally enlarged or compressed depending on the desired
performance of the min-max novelty detector.
4-1
9807070
Novelty Filter for Short Detection
The nearest neighbor novelty detector allows for more general data topology. Here, minimum
Euclidean distances are found between each point and its closest neighbor. The distance
proportional to the maximum of these distances is then used as a decision parameter. Every
incoming point is compared to every point in the healthy training set. If the new point is at a
greater distance from each of the healthy points than the decision parameter, it is declared to be
novel.
Novelty detection can be couched in the paradigm of hypothesis testing.
H0: status quo (no winding shorts)

H1: a winding short is present.
The binary hypothesis test considers a given hypothesis H0 that is to be proved, versus an
alternative hypothesis H1. In our case, H0 is the hypothesis that the rotor is healthy. H1 is the
alternative hypothesis that the rotor has shorts. In marking a decision, two types of error can
occur. The false alarm probability is
α = Probability[ H1 is announced given H0 is true ]

= Probability[ a short is erroneously announced ].
and the detection probability is
β = Probability[ H1 is announced given H1 is true ]

= Probability[ a short is correctly announced ].
For problems where novelty detection is used, the detection probability can not generally be
estimated. This is due to the unavailability of shorted winding signatures. In other words, we
have no data corresponding to when “H1 is true”. The false alarm probability of the novelty
detector, on the other hand, can be straightforwardly evaluated. After the novelty boundary is
established using status quo training data, additional status quo test data is collected. The
percentage of times a short is announced from the test data is an estimate of the false alarm rate.
The law of large numbers assures convergence as the cardinality of the test set increases.
There is an inherent tradeoff between the false alarm rate, α, and detection rate, β. As one
increases monotonically, so does the other. Each of the novelty detectors described has a
parameter that allows tuning of this tradeoff. In the spherical and ellipsoidal cases, it is the
choice of radius and Mahalanobis distance respectively. As the radius increases, both the
detection and false alarm probabilities increase monotonically. For the nearest neighbor
detector, the tradeoff is similarly determined by the ball drawn around each training data point.
The α and β parameters for the min-max detector are tuned by the choice of padding about the
detection box region.
The only way of controlling the performance in novelty detection is to define an acceptable false
alarm rate not to be exceeded by the training set. An alternate method for setting the threshold,
not considered here, requires the finding of threshold from the allowable false alarm probability
from parametrical estimators. This method requires an assumption about the probability
distribution of the data.
4-2
9807070
Other types of novelty detection have been suggested. One often referred to as the novelty filter
[10], compares the orthogonal complement of the linear space spanned by the training set to a
threshold. In geometrical terms, this method tries to fit a hyperplane to the data. Only the
orthogonal distance to the hyperplane is considered. Such an approach is useful only when the
healthy data are known to lie in a hyperplane. Other approaches assume that the healthy data lie
in clusters. These were also not applied to our problem since the healthy data appeared in one
large cluster. One example is use of radial basis neural networks [22] as a method of a non-
parametric estimation of the data’s a priori distribution, which has also been applied to novelty
detection [17]. A statistical semi-parametric estimation technique defining several
hyperellipsoidal clusters is described in [18] and applied, extended, to novelty detection in [19].
This is an extension of the elliptical boundary detection to multiple clusters. A robust statistical
method for finding elliptical clusters is defined in [20]. Nonlinear statistical estimation is
applied in [21] where a neural network is trained to recognize a mapping of any given probability
distribution to an uncorrelated Gaussian distribution. This is done with an information
preservation criterion, and a simple spherical boundary detection is then applied. Other methods
exist such as ART clustering techniques, which automatically defines new clusters as something
new is observed.
Which novelty detection approach works best? The answer depends on the structure of the
healthy data [25]. In our case, the elliptical novelty detector gave the best performance on the
healthy data collected for our rotor under load. In the test of the performance of the elliptical
filter, there was rare availability of shorted winding signatures. This allowed testing for the short
detecting capability of the novelty detectors.
4.1 Computation of Detection Threshold
The tradeoff between the detection probability and the false alarm probability is determined by
choice of a threshold.
The first operation performed on the signature signals is the removal of the mean. Since the
signature signals are sampled and converted to an array of digital numbers, each signal can be
considered a discrete vector or a point in the signal space. The average signal is computed by
summing all healthy rotor signatures as vectors and dividing each component by the number of
healthy signature signals used in the sum. The average vector, called the prototype signature, is
subtracted from each signature signal to translate the signature signals toward the signal space
origin.
The signal space is then partitioned into two regions: one for healthy rotors and one for faulted
rotors. Due to the stochastic nature of the signature signals, the region corresponding to healthy
rotors will extend away from the origin. A surface separating the two regions must be defined to
enclose the healthy region. Any translated signature signals outside the detection surface is
assumed to represent a rotor with a fault.
The simplest detection surface is a hypersphere. The largest Euclidean length of the translated
healthy signature signals is used as the threshold. The region for healthy rotors is inside the
resulting hypersphere. Any signature signal outside the hypersphere will represent a faulted
rotor. The detector's decision is made according to
4-3
9807070
Fault = True if x − x > T

(4.1)
= False if x − x < T
where x is a signature signal, x is the prototype vector and T is the detection threshold or radius
of the hypersphere. This method could potentially produce a detection algorithm with a high rate
of missed detections. For example, if the translated healthy signatures all lie on a line, the sphere
will enclose all the points but will have a great deal of space with no healthy signature signals
nearby. A signature signal for a faulted rotor may lie within the hypersphere but not on the line
and thus no fault would be detected.
A more accurate detection boundary can be defined using hyperellipses. The hyperellipse can be
oriented in any direction and thus would not contain large empty spaces unless the healthy region
consisted of multiple disjoint subregions. A signature matrix X is formed by using the healthy
translated signature signals as the columns. The correlation matrix C is formed by
C = X XT (4.2)
The eigenvectors of the correlation matrix C define the principle axes of the hyperellipse. Each
translated healthy signature signal is rotated into a new space by vector multiplication by the
eigenvectors.
The use of the full signature matrix results in a large computational burden. The signature
signals obtained for laboratory experiments have typically contained a minimum of 500 samples.
The correlation matrix would then contain 5002 = 250,000 elements. Computing the
eigenvectors of such a large matrix can be difficult especially on the small personal computers
used for data acquisition. For this reason, a neural network feature extraction method is used to
reduce the dimension of the signature signals before construction of the correlation matrix
[12-15].
The feedforward neural network is trained to reproduce the input vectors at the output.
However, a hidden layer containing a small number of nodes is used. If a linear network is used,
the resulting operation can be shown to correspond to a projection onto a linear subspace. The
dimension of the subspace corresponds to the number of nodes in the hidden layer. The
activation values of the hidden nodes then correspond to the extracted features of the signature
signals. To improve the performance of the linear network, the inputs are scaled to be within the
interval [0, 1].
If the signature space is imagined to be three dimensional, the projection performed by the neural
network can be viewed as a projection onto a plane. The hyperellipse detection boundary then
lies on the plane. This is illustrated in Figure 4-1.
4-4
9807070
Measured
Signature
Signal
Toc Decision
Hyperellipse
Projection
Figure 4-1
Simplified view of the projection of a signature onto the hyperellipse. The measured
signature is shown to be projected onto the hyperellipse but to be significantly removed
from the projection plane
The possibility the projection of a shorted winding signature may be inside the resulting
hyperellipse must be addressed. To resolve this problem, the magnitude of the orthogonal
complement is computed for each healthy signature. Since the training of the network will not
be prefect, all signals will have a small orthogonal complement. The orthogonal complement
can be computed by subtracting the network output from the input vector. The maximum
Euclidean vector magnitude of the orthogonal complements of all healthy signatures can be
stored to identify which vectors are potentially outside the detection surface. Let the maximum
magnitude be given by Toc.
The threshold calculation process now consists of translated healthy signatures which have been
reduced to a smaller dimension. Since the signatures are scaled between zero and one and may
not be centered about the origin, a new prototype is calculated for the reduced data and removed
from all healthy, reduced signatures. The correlation matrix can then be constructed and the
eigenvectors, E, easily calculated. The eigenvectors correspond to the axes of the hyperellipse.
By forming the dot product between the eigenvectors and the signatures, the ellipse can be
oriented along the coordinate axis making ellipse calculations much easier.
The next step involves calculation of the size of the ellipse or the actual detection boundary. A
hyperellipse with minimum volume which encloses all healthy signatures is desired. This is
done using the following procedure. First, all ellipse coordinates are set to the distance to the
farthest healthy signature forming a hypersphere. Then, for each coordinate, the minimum size
enclosing all signatures is computed by
 
 
 
 
aj = &
max 
x 

xj
N −1   2

 (4.3)
x
 1 − ∑  i  
 i =0  ai  
 i≠ j 
 
4-5
9807070
where N is the number of healthy signatures, aj is coordinate j of the hyperellipse, and xj is

component j of the signature. The maximum is taken over all healthy signatures. The
hyperellipse coordinates must be less than the maximum magnitude used to initialize the
coordinates and each coordinate is independent of the others. Thus, the above procedure will
result in the desired minimum volume hyperellipse.
The test to determine if a signature is within the hyperellipse is another simple calculation. The
value B given by
N −1 
2
x
B = ∑  i  (4.4)
i = 0  ai 
The test determining whether the signature is inside or outside the detection boundary is
Fault = True if B > 1

(4.5)
Fault = False if B < 1
4.2 Mahalanobis Distance
The principal component analysis technique can be used to maximize the data separation [24].
Assume that, for Equation (4.2), C has N samples in a p –dimensional space. The principal
components are found by means of orthonormal eigenvector decomposition, so that
C vi = λi vi
1 for i = j (4.6)
where viT vi = 
 0 for i ≠ j
where vi is the eigenvector corresponding to the eigenvalue λi for data index i.
The rank of (X) is
Rank ( X ) = min[ N − 1, p ]
To ensure a full rank matrix, enough data should be taken so that N − 1 ≥ p . The largest
eigenvalues and eigenvectors determine the principal components of the data. To achieve this,
all eigenvalues are ordered and Mahalanobis distance is computed for each data set. Define
∑ λ (v (x )) (v (x ))
m
1 T
di = T
j i −x T
j i −x (4.7)
j =1 j
,
4-6
9807070
where d i is the Mahalanobis distance, and m is the cut number of the data set. The value of d i is
scaled
∑ λ (v (x )) (v (x ))
m
1 T
dl = T
j l −x T
j l −x k (4.9)
j =1 j
where
k = max (d i )
i =1... N
For “healthy” rotor, all d l ’s should be < 1. A value greater than 1 indicates a short. In a noisy
environment, a threshold should be set greater than 1.
4.3 Lab Evaluation
The lab prototype was used to test the novelty of the signature signals. A set of signatures was
gathered at the operating stages corresponding to
1. stopped rotor,
2. turning-gear speed, and
3. rotors running at a full speed.
The stopped rotor situation is similar to off-line testing except that the signature signal is
obtained through the brushes. The rotor is in turning-gear when it is rotating very slowly - in our
case at 30-60 rpm. The fast rotating rotor ran at the synchronous speed, 1800 rpm, to simulate a
four-pole turbine-generator in full operation.
Three different novelty detectors were constructed corresponding to three different speeds. This
approach was motivated by the observation that the signature signal changed significantly for
varying speeds as shown in Figure 4-2. The full-speed rotation provided by the external machine
strongly affects the signature waveform through slip-ring noise and general impedance change in
the windings at the higher frequencies. However, the stopped rotor and turning gear signatures
appear similar. Due to the similarities, combining the stopped rotor and turning gear ratio
novelty detectors into a single detector is a possibility in future studies.
4-7
9807070
0.8 Stopped Rotor

Turning−Gear
Full Speed
0.6
0.4
0.2
−0.2
−0.4
−0.6
−0.8
0 0.5 1 1.5
Time [s] −5
x 10
Figure 4-2
Typical signature signals for different rotor speeds under no load
A given novelty detector was expected to perform over a large range of loads. The variation of
the signature signals under varying loads at the same speed, as shown in Figure 4-3, was not as
significant. The two dimensional plane for both cases is obtained from the two largest principal
components of the data. Although there is minor drift, the data in both cases clusters nearly
independent of the load. For a given operating speed, this is the case generally. Thus, the
shorted winding novelty detector need not be parameterized by the load.
Five excitation levels between 0 and 1.6 A are used in each training set. Each training set
consists of 1000 signatures taken in sets of 50 each at different times during one day. The
novelty detectors designed from this data are tested against five “healthy” test sets taken over a
period of 3 days. Each has 200 signatures, so a total of 1000 signatures of healthy data aretested.
Shorts are induced around all the poles of the rotor, three for each inner winding and two for
each outer. This is also done for all the operating conditions described above. For each position
of shorted turn and voltage level, 10 samples are taken for a total of 1000 signatures for each
speed. Figure 4-4 shows examples of some shorted-turn signatures compared to a typical non-
shorted one. Even visually the variation is significant enough to be detected.
The feature vector used is the time signature waveform, sampled at 20 MHz and decimated (i.e.
downsampled) by 4 to reduce computation. Experience shows decimation does not effect
detection performance. Decimation leaves 75 points in the signature signal vector. No
additional feature extraction method is applied.
4-8
9807070
0.2
0.1
PC #2 0
−0.1
−0.2
−0.3 0.0 A
0.4 A
0.8 A
−0.4
1.2 A
1.6 A
−0.5
−0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
PC #1
0.2
0.1
0
PC #2
−0.1
0.0 A
0.4 A
0.8 A
1.2 A
−0.2 1.6 A
0 0.1 0.2 0.3

PC #1
Figure 4-3
Signature signals of healthy rotor obtained from a rotor operating at full speed (top) and
stopped (bottom) under different excitations levels
4-9
9807070
Non−shorted
0.8
Short in Pole 1
Short in Pole 2
0.6
Short in Pole 3
Short in Pole 4
0.4
0.2
−0.2
−0.4
−0.6
−0.8
−1
0 0.5 1 1.5
Time [s] −5
x 10
Figure 4-4
Signature signals for a stopped rotor with shorted turns
Table 4-1 shows the detection results obtained for the different operating stages. The false alarm
rate, α, is averaged over the five acquisition periods and the detection probability, β, is from
signatures obtained over a 24 hour period. The best performance comes from the elliptical
novelty detector.
Table 4-1
Detection results for voltage-excited running rotor
Stopped Turning-Gear Full Speed

Rotor Rotor
Detection α% β% α% β% α% β%
Method
Spherical 0.0 83.8 0.0 85.3 0.0 65.0
Min-Max 63.6 100 32.9 100 0.8 92.8
Nearest 56.2 100 0.0 100 0.1 79.3

Neighbor
Elliptical 0.0 100 0.0 100 0.4 91.0
4-10
9807070
Figure 4-5 shows plots of α (left hand curves) and 1–β (right hand curves) for varying
thresholds. All methods give good results for the proper threshold. In practice, however, the
miss probability cannot be estimated since data for shorted windings is not available. This plot,
shown for the rotor at full speed, is typical. The false alarm rate can be read from the curve on
the left and the miss rate from the curve on the right. For novelty detection, the miss rate can
not, in general, be measured as it is in this case. Thus, although there exists a threshold
corresponding to low false alarm and miss rate for each detector, it can not, in general, be set by
observing both of these curves.
0.8
0.6
Error Rate
0.4
0.2 Min−max Boundary

Spherical Boundary
Nearest Neighbor
Elliptical Boundary
0 −1 0 1
10 10 10
Threshold
Figure 4-5
False alarm, and miss rate for different thresholds
4-11
9807070
9807070
5
NEURAL NETWORK FOR SHORT LOCALIZATION
Since introduction of the multi-layer perceptron neural network (NN) technology and efficient
training algorithms, NN's have proven useful in several power applications. NN's, when
properly designed and trained, can synthesize a useful nonlinear mapping between input and
output patterns. This ability proves to be a key property for short turn detection and localization
[7-8].
5.1 Fuzzified Neural Network
The shape of the signature signal of the twin signal technique is used by the NN to perform two
functions: 1) detect the existence of a shorted turn; and 2) localize the short. A high frequency
sampling device is used to ensure that the entire signature signal is captured. This results in a
vector whose high dimension cannot be easily processed by a neural network. Hence, the
cardinality of the training data must be reduced without destroying the data's information
content. Figure 5-1 outlines the general procedure for the NN short turn detection and
localization. The training data acquired by the setup for acquisition of the signature signal is
used for neural network training. Extracted features from the signature signal are used as inputs
to train a standard feed forward layered perceptron artificial neural network with outputs coded
as fuzzy membership functions. The network output is defuzzified to provide a crisp number to
identify the short location.
6DPSOHG 'DWD
0HDVXUHPHQWV 3UHSURFHVVLQJ
)HDWXUH
([WUDFWLRQ
7UDLQLQJ 1HXUDO 1HWZRUN )X]]LILHG

1HXURQV
7HVWLQJ
'HIX]]LILFDWLRQ
5HVXOWV
Figure 5-1
Short turn localization procedure
5-1
9807070
Neural Network for Short Localization
The location of a shorted winding is coded into a number of fuzzy membership functions
determined by the desired resolution of the short location. For example, a field winding may be
divided into coils and the coils divided into turns. In this study, six membership functions are
used. The number of output neurons of the NN is the same as the number of the fuzzy
membership functions. This is illustrated in Figure 5-2. The turbogenerator used to test this
concept has 14 coils with 17 turns in each coil for a total of 238 turns. These turns are
partitioned into six groups. Each output neuron corresponds to the value of the corresponding
membership function. For a short at turn #75, the membership (neural network output) is [0 0.4
0.6 0 0 0]T.
Membership
(Fuzzified) 1
Neuron 2 0.6
Neuron 1 0.4 Location
0 1 2 3 4 5 (Target)
Turn #75
Turn Turn Turn Turn Turn
# 1-47 # 48-94 # 95-141 # 142-188 #189-238
Neural Network
Output Neurons
Figure 5-2
Membership function for the fuzzy mappings
During testing, the output of the neural net is defuzzified where each membership function is
weighted by the state of the corresponding output. The weighted membership functions are then
added and the center of mass (first moment) of the sum is the short location. If, for example,
each membership function is of identical shape and has a center of mass Ci, then the defuzzified
output is given by
∑ β i Ci
short location = i (5.1)
∑ βi
i
where βi is the output of neuron i.
This detection method was tested in 1996 in the Southern California Edison Company facilities
on a 60 MVA 2-pole turbogenerator. The generator has 14 coils with 17 turns in each coil. The
shorts between windings were intentionally introduced to verify the proposed technique. The
loss of one turn reduces the ampere-turns of that pole by about 0.85%.
5-2
9807070
Turn #7
2
1.5
1
Voltage
0.5
0
-0.5
-1
0 20 40 60 80 100
Time (microsecond)
Turn #49
2
1.5
1
Voltage
0.5
0
-0.5
-1
0 20 40 60 80 100
Time (microsecond)
Turn #209
2
1.5
1
Voltage
0.5
0
-0.5
-1
0 20 40 60 80 100
Time (microsecond)
Figure 5-3
Samples of signature signals
To train and test the NN, temporary shorts are introduced between adjacent windings. Two
simultaneous twin signals are then injected from both sides of the field winding. The difference
between the two, the signature signal, is used to detect and localize the shorted turn. Examples
of several sampled waveforms for these windings are shown in Figure 5-3. The horizontal axis
represents the time in microseconds and the vertical axis is the magnitude of the signature signals
in volts. The signature signal is the difference A-B.
The signature signals are sampled at 5 MHz and a total of 500 samples are collected. If the
entire signal is used to train the neural net, the NN suffers from scaling problems and the curse of
dimensionality. As discussed, feature extraction must be used to capture the information content
of the signal and reduce the dimension of the sampled signal vector. Some feature extraction
methods are based on mathematical techniques [12-15]. Others are based on engineering
judgment and heuristics.
Figure 5-4 illustrates a signature signal. Dispersion causes the signature signal to have large
differences for different short locations near the signal's source. Therefore, the signature signal
is divided into two sections: initial and extended. In the initial section, the waveform is
5-3
9807070
partitioned into four intervals of 18 microseconds each. The extended section is composed of
four intervals of 36 microseconds and five intervals of 72 microseconds. The area of each
interval is given by
ti
Areai = ∫ v(t) dt (5.4)
t i -1
where Areai is the area of segment i, ti is the end time of segment i, and v(t) is the voltage
waveform. Since the signal is sampled, the area is merely taken as the discrete voltage sample
multiplied by the width of segment i. A total of thirteen areas are obtained for each signature
signal. These areas are the features used as the input to the NN.
v(t)
Area
i
t t t t t
0 1 2 3 4 t5 t6 t7 t 8 t
9 t
10 t
11 t
12 t
13
18 us 36 us 72 us
Figure 5-4
Features extracted from signature signal
A total of 67 training patterns were collected by shorting adjacent turns at several locations
within the field winding. A NN with one hidden layer, thirteen input neurons, four hidden
neurons, and six output neurons is used. This architecture gives a lower test error than other
architectures tried. The network is trained by using standard back-error propagation.
After training, the NN is tested for 60 short locations taken at random points. None of the test
data is used during training. The test results are listed in the Table 5-1. As seen in the table, the
proposed technique is both highly accurate and very robust. In all test cases, the coil with
shorted turns was accurately identified. Moreover, the shorted turns were all located to within a
few turns of the actual short location.
The technique for locating shorted windings requires the training of a NN with input/output
pattern pairs. Thus, the ability to introduce shorts for acquiring training data is essential. The
expense of dismantling and reassembling a large turbogenerator prevents the introduction of
shorts for training sample acquisition when shorts are to be detected in an operational rotor.
Therefore, an alternate technique is required to detect shorted turns without the need to introduce
shorts into the field windings.
5-4
9807070
Table 5-1
Field test results
NN Structure
Input Neurons 13
Hidden Neurons 4
Output Neurons 6
Patterns
Training Patterns 67
Testing Patterns 60
Test Results
Percentage of Accurate 100%

Identification of Coil
Maximum error in short ± 7 turns (± 3%)

localization
5.2 Considerations for Operational Generators
The detection of shorted turns in operational rotors can be accomplished using the traveling wave
technique and novelty detection. The detection of faults for a rotor which is in operation is made
much more difficult due to the presence of noise caused by the brushes, the mechanical
variations during rotation, and the presence of the excitation source in the rotor circuit.
The brushes of a turbogenerator connect the excitation source to the slip rings and supply power
to the rotor. The motion of the brushes and the current flowing through the brushes causes
arcing. The arcs produce high-energy spikes that appear as noise in the signal. The noise will be
measured by the signature signal acquisition hardware and will corrupt the signal. A study of
brush noise was performed on a small laboratory machine and the results are shown below.
The rotation of the rotor causes the windings to move and shift as the rotor turns. The signature
signal is dependent on the physical characteristics of the rotor windings. The signature signal
will thus not be identical when measured at different points in the rotational cycle.
The excitation source provides another path for the pulses of the signature signal acquisition
system. The added path greatly changes the reflected signals and the signature signal. As long
as the excitation windings are symmetrical from the perspective of the injected pulses, the ability
to detect faults, though, will not be hindered.
5-5
9807070
All of the modifying effects described prevent the use of training examples collected while the
rotor is dismantled to detect shorted turns while the rotor is operating. Since inducing shorted
turns in an operational rotor is expensive, novelty detection must be used to assess the presence
of shorted turns.
The acquisition of signature signals for operational rotors is similar to the acquisition of
signature signals for dismantled rotors (See Figure 2.4.) The addition of blocking capacitors in
series with each pulse generation path is required to protect the measurement circuits from the
high excitation source voltage present on the brushes during operation. The excitation source
required by most turbogenerators supplies a high DC voltage that can easily be blocked by the
capacitors. The pulses are not significantly affected by the capacitors and thus measurement of
the signature signals is not seriously affected.
Since the signature signals are corrupted by noise, the signature signals are averaged over many
collected waveforms. Averaging of the signals will reduce noise with zero mean such as brush
noise. The number of signals averaged depends on the standard deviation of the noise and the
degree to which the noise is to be removed. For noise outside the bandwidth of the signature
signals, a lowpass frequency selective filter is used. Different rotors will produce signature
signals with different frequency content. Therefore, the signature signals and the undesired noise
frequency spectrums must be analyzed for each rotor. The spectrum of the noise is easy to
obtain using the signature signal acquisition hardware. The pulses should be disabled and many
samples of the noise signal gathered. The average power spectral density can then be computed.
5.2.1 Brush noise characterization
Brush noise effects signals used to detect rotor winding faults while the machine is operating.
The noise on the brushes operating at full load was measured and collected using a Tektronix
TDS540 oscilloscope and a GPIB interface to a personal computer. The power spectrum of the
stored noise data is computed using the following procedure: (1) The mean of each signal is
removed. (2) The autocorrelation function is estimated by computing the magnitude squared of
the Fourier transform of the signal and inverse transforming. The inverse transform is then
scaled to produce an unbiased estimate of the autocorrelation function. (3) The estimated
autocorrelation functions of several (usually 10) collected signals are averaged and windowed to
remove edge transitions. A Hanning window was used in this study. (4) The magnitude of the
Fourier transform of the windowed autocorrelation average was computed to produce the
estimated power spectral density of the noise signal.
Figure 5-5 shows the measured signal with the machine rotating at rated speed (1800 rpm) under
rated load (field excitation of 90 volts). The slow oscillation is the full wave rectified power
supply ripple at 120 Hz. The noise at the sampling frequency used for this measurement is not
high enough to prevent aliasing. The sampling frequency at which aliasing did not occur was
determined by progressively increasing the sampling rate until the power spectrum did not
change. The step increases in sampling rate require that the data be collected using 25M samples
per second to avoid aliasing.
5-6
9807070
0.5
Signal Amplitude (volts)
0.5
0 0.005 0.01 0.015 0.02
Time (sec)
Figure 5-5
Brush noise measurement showing noise spikes and power supply ripple
1
Signal Amplitude (volts)
0.5
0.5
1
0 0.001 0.002 0.003 0.004 0.005
Time (sec)
Figure 5-6
Brush noise measurement showing the periodic nature of the noise spikes
One further observation about the data in Figure 5-5 is the periodic noise spikes. These spikes
occur at approximately 1500 Hz. The periodic nature is more evident in the expanded signal
shown in Figure 5-6. The analysis of the power spectrum of the brush noise in general is
hindered by this artifact. The oscilloscope trigger is required to capture the noise spikes within
the displayed trace. However, the spikes are then over represented in the data. One remedy is
removal of brush noise by averaging many sweeps of the data. If the oscilloscope trigger is used
to catch the spikes, though, the average is not zero since the spikes have a similar shape from one
sweep to the next. But if the oscilloscope trigger is not used, the probability of collecting a spike
becomes small. The ability to remove the brush noise was thus verified visually by noting the
absence of spikes flashing across the oscilloscope screen. Data is collected and analyzed by
repeatedly triggering the oscilloscope at random points in an attempt to capture the noise
character. The resulting power spectrum is shown in Figure 5-7 for averaging of 100 sweeps
before collection. Note that, although the probability of capturing a noise spike in the
oscilloscope trace is very low, the probability that a fault detection pulse will overlap with a
noise spike is much higher. The time between oscilloscope samples is very small but the time
between sweeps quite large. As described below, the noise spike frequency is relatively close to
the pulse frequency of the fault detection system.
5-7
9807070
0.001
Power
0.001
0
0 1 2 3 4 5 6 7
Frequency (MHz)
Figure 5-7
Power spectrum of brush noise where each measurement was the average of 100
oscilloscope sweeps
1
Power
0.5
0
0 1 2 3 4 5 6 7
Frequency (MHz)
Figure 5-8
Power spectrum of brush noise
Figure 5-8 shows the power spectrum of the voltage across the brushes at rated speed and field
excitation using the oscilloscope trigger to capture the noise spikes. Averaging was not used
during this measurement. Note the difference in the scales between Figure 5-7 and Figure 5-8.
The frequency content appears to be roughly the same except the power scale is greatly
expanded. The noise has been drastically reduced by the averaging process.
5.2.2 Field Verification
Signature signals were recorded for two rotors of small generators in various modes of operation.
Signature signals for the first machine are taken both while the rotor is motionless and rotating at
various speeds, with and without the brushes contacting the rings, and with and without the
excitation source connected to the rotor. Signature signals for the second machine are only
collected with the brushes removed from the slip rings and the rotor at standstill. In both cases
the instrumentation circuit is not grounded directly at the rotor. The signature signals are
recorded at 250M samples per second with a 20 MHz low pass filter. The signature signals
consist of 1000 samples of the reflected difference signal. Each signature signal is the average of
5-8
9807070
50 individual signals to remove the effects of noise. The averaging was performed by the
TDS540 oscilloscope.
The second experiment was conducted with the brushes in contact with the slip rings and the
signature signal measured through the brushes. Figure 5-9 shows the signature signal when
connected directly to the slip rings (solid line) and through the brushes (dotted line). The
signature signal obtained through the brushes has a significantly different shape from the
signature signal applied directly to the slip rings. Though not shown in the figure, connection
variations, as expected, significantly altered the signature signal.
0.5
Amplitude (Volts)
0.5
1 7 6 6 6 6 6 6 6
0 5 10 1 10 1.5 10 2 10 2.5 10 3 10 3.5 10 4 10
Time (sec)
Figure 5-9
Plot of signature signal taken directly from the slip rings (solid line) and the signature
signal taken through the brushes (dotted line)
The next experiment consists of connecting the excitation source to the rotor. In several stages,
the rotor is rotated at full speed with the excitation connected thus providing power to the rotor.
The amount of power applied to the rotor is proportional to rotation speed. Figure 5-10 shows
plots of the signature signal at several different speeds. The solid line represents no rotation, the
dotted line represents a very slow rolling rate, the dashed line represents 1800 rpm, and the dash-
dot line represents 3600 rpm. All of the signature signals are seen to be similar especially close
to the beginning. As the rotation speed increased, the signature signals became more unstable, as
if being modulated by a low frequency signal. The same effect is noticed in a laboratory
experiment where the low frequency modulation at 120 Hz results in AC ripple on top of the DC
excitation source.
Signature signals from a second machine were recorded by connecting the measurement leads
directly to the slip rings. The rotor was not rotating during this test. Figure 5-11 shows the
signature signals for both machines under this condition. The two signature signals are very
different implying that the machines are either different or that one machine has shorted turns.
The resistance to ground for the second machine, measured before the experiment began, was
found to be much lower than the resistance for the first machine. It is not clear that the resistance
to ground should have such an impact and further studies are indicated.
5-9
9807070
0.5
Amplitude (volts)
0.5
1.5 7 6 6 6 6 6 6 6
0 5 10 1 10 1.5 10 2 10 2.5 10 3 10 3.5 10 4 10
Time (sec)
Figure 5-10
Signature signals for various rotation rates
The signature signals collected during this experiment demonstrate that detection of shorted rotor
windings may indeed be possible while the rotor is in operation. Several tests are required to
improve the detection accuracy. Since the signature signals for the two similar machines are
different, a test must be conducted on two machines which are known to have identical rotors
and excitation sources and known not to have shorted rotor windings. The resulting signature
signals will verify whether the detection process requires signature signals for all machines
before shorted windings can be detected or if a single signature signal for a given machine type
will suffice.
0.5
Amplitude (Volts)
0.5
1 7 6 6 6 6 6 6 6
0 5 10 1 10 1.5 10 2 10 2.5 10 3 10 3.5 10 4 10
Time (sec)
Figure 5-11
Signature signals for two different machines. The measurement leads were connected
directly to the slip rings
5-10
9807070
6
LABORATORY TEST
The detection and localization methods developed in this project were tested in the CIA lab by
using the prototype described in Section 2 of this report. A set of signatures was gathered at
several different operating speeds and excitation currents. The operating speeds are standstill,
turning gear and full speed. The turning-gear speed is 30-60 rpm. The full-speed is 1800 rpm to
emulate a 4-pole speed.
Two novelty detectors were built: one for standstill and turning gear speed, and the other for full
speed. The reason for using two detectors is the change in the signature signal shape at full
speed. Figure 6-1 and 6-2 shows the signature signals for the twin and in-out methods
respectively.
To observe the effects of shorted windings, 27 shorts were made across adjacent windings
covering all slots, poles and coils. Several of these shorts are near the center and the edges of the
coil. A total of 600 signatures of healthy rotor data were collected for all three speeds. For each
short, 50 samples are collected. A total of 5 excitation current levels were used at full speed.
Figures 6-3 and 6-4 show examples of signature signals for various shorts.
0.5
Stopped rotor
Turning gear
0.4 Full speed
0.3
0.2
0.1
-0.1
-0.2
-0.3
-0.4
-0.5
0 0.5 1 1.5 2 2.5 3
Time (s) -5
x 10
Figure 6-1
Typical Signatures for Different Speeds (Twin Signal)
6-1
9807070
Laboratory Test
Stopped rotor
Turning gear
Full speed
0.5
-0.5
0 0.5 1 1.5 2 2.5 3
Time (s) -5
x 10
Figure 6-2
Typical Signatures for Different Speeds (In-Out Signal)
0.6
Non-shorted
Short in Pole 1
0.4
Short in Pole 2
Short in Pole 3
Short in Pole 4
0.2
Short in Pole 5
-0.2
-0.4
-0.6
-0.8
0 0.5 1 1.5 2 2.5 3
Time (s) -5
x 10
Figure 6-3
Shorted Turn Signatures for Stopped Rotor (Twin Signal)
6-2
9807070
Laboratory Test
1
Non-shorted
Short in Pole 1
Short in Pole 2
0.8 Short in Pole 3
Short in Pole 4
Short in Pole 5
0.6
0.4
0.2
-0.2
-0.4
-0.6
0 0.5 1 1.5 2 2.5 3
Time (s) -5
x 10
Figure 6-4
Shorted Turn Signatures for Stopped Rotor (In-out Signal)
6.1 Short Detection
Table 6-1 shows the false alarm (α) and detection rate (β) for the twin and in-out methods.
Remarkably, 100% detection results is achieved in all cases. Because the shorts did not cause
asymmetry in the reflected signals, the detection rate is perfect. Keep in mind that the in-out
technique requires matching impedance at one of the winding’s ends. Once the matching
impedance is determined, a short may conceivably occur in such a way that the signals are
symmetrical. In such a case, the in-out technique may fail. Moreover, the twin-signal is not
accurate when the short is near the center of the winding. The two methods are thus
complimentary. In practice, both should be used.
Table 6-1
Detection Results for Running Rotor Using Twin Signal and In-out Signal
Standstill Turning Gear Full Speed (No Full Speed (With-

Excitation) Excitation)
Detection α β α β α β α β
Method % % % % % % % %
Twin signal 0.0 100 0.0 100 0.0 100 0.0 100
In-out signal 0.0 100 0.0 100 0.0 100 0.0 100
6-3
9807070
Laboratory Test
0.9
0.8 The false alarm rate
0.7
0.6
Error Rate
0.5
0.4 The missed detection rate
0.3
0.2
Standstill
0.1 Turning Gear
Full speed
0
0 10 20 30 40 50 60 70 80 90 100
Threshold
Figure 6-5
ROC Curves for the Running Rotor without DC-Excitation (Twin Signal)
Several empirical ROC (receiving operator characteristic) curves were computed for various
test cases. The results are shown in Figures 6-5 through 6-8. In these figures, α is the falling
curve, and 1-β is the rising curve. The ROC curves show that each method can give superb
results for a reasonable choice of threshold.
1
0.9
The false alarm rate

0.8
0.7
0.6
Error Rate
0.5
The missed detection rate
0.4
0.3
0.2
Standstill
0.1 Turning Gear
Full speed
0
0 10 20 30 40 50 60 70 80 90 100
Threshold
Figure 6-6
ROC Curves for the Running Rotor without DC-Excitation (In-out Signal)
6-4
9807070
Laboratory Test
0.9
0.8
0.7
0.6
Error Rate
0.5

0.4
0.3
0.2
Current 1
Current 2
0.1 Current 3
Current 4
Current 5
0
0 5 10 15 20 25 30
Threshold
Figure 6-7
ROC Curves for DC-Excitation (Twin Signal)
0.9
0.8
0.7
0.6
Error Rate
0.5
0.4
0.3
0.2
Current 1
Current 2
0.1 Current 3
Current 4
Current 5
0
0 5 10 15 20 25 30
Threshold
Figure 6-8
ROC Curves for the DC-Excitation (in-out Signal)
6-5
9807070
Laboratory Test
The twin and the in-out signal sensing methods are effective techniques for short detection. They
can both be applied on-line to operational machinery. For the two techniques, the basic rotor and
winding configuration is the same. As explained in Section 2, the setup for both methods is
composed of two parts: signal generation and signal detection. The signal generator creates two
high-frequency pulses. For the twin signal sensing method, the signals are injected into each side
of the winding under test and the difference between the two reflected signals is measured. For
the in-out signal sensing, the high-frequency pulses are injected and an impedance match is
connected to one of the ends. In both configurations, the signature signal is related to the
configuration of the test rotor. If shorts occur, changes in the rotor signatures are to be expected.
Both the twin and the in-out signal methods rely on deviation from symmetry. For the twin
signal method, if a short occurs near the center of the windings the signature signal does not
readily reveal the short. Hence the detection is difficult. For the in-out signal method, if the
short occurs at a point where the fixed impedance match configuration creates symmetry, the
method can fail. To avoid this problem, both methods should be used in practice
As given in Equation (4.9), the novelty distance d i should be made as large as possible to ensure
a separation between the healthy and short data. To compare the two techniques, let us examine
Figures 6-9 and 6-10. Figure 6-9 is for the twin-signal method and Figure 6-10 is for the in-out
method. In these figures, the data at the bottom represents the healthy data, and the top
represents the data with shorted turns. As seen in the figures, both methods perform well, but the
twin signal provides a better overall separation.
Figures 6-11 and 6-12 show the case when the short is near the center. In this case, the in-out
method provides a much better separation as compared to the twin-signal.
250
200
150
100
50
0
0 50 100 150 200 250
Figure 6-9
The Elliptic Boundary Technique using Twin-Signal method
6-6
9807070
Laboratory Test
70
60
50
40
30
20
10
0
0 50 100 150 200 250
Figure 6-10
The Elliptic Boundary Technique using in-out method
14
12
10
0
0 50 100 150 200 250
Figure 6-11
The Elliptic Boundary Technique using Twin-Signal method
6-7
9807070
Laboratory Test
30
25
20
15
10
0
0 50 100 150 200 250
Figure 6-12
The Elliptic Boundary Technique using in-out method
6.2 Short Localization
Several metrics can be used to localize a short. One metric is described in Section 5 where the
entire signature signal is used to train NN at various short locations. Here, the resistance
between the short location and one end of the rotor is used as a localization metric. The NN is
trained to predict this resistance.
Feature extraction is used to reduce the input number of the neural network in order to avoid the
curse of dimensionality [12-15]. The signature signal is sampled at a high rate that, due to the
resulting high dimensionality, can not be used as input to the NN. Only 5 features are extracted
from the signature signals.
A total of 450 signatures were collected for several shorts - 70% of them are used for training,
and the rest for testing. A NN with two hidden layers, seven nodes each, five input neurons, and
one output neuron was used. This architecture gave a lower test error than other architectures
tried. After training, the neural net was tested for 135 shorted data. None of the test data was
used during training. The training epoch is 5000. The error margin is 0.1. The result is shown in
Table 6-2. Here, the “average RMS error” is the average of the RMS error over all patterns in
the training and testing data. The “maximum RMS error” is the pattern with the largest RMS
error among all output neurons. The “maximum error” is the error for a particular pattern with
the largest output error among all output neurons and patterns. For networks with a single output
neuron, the max error and max RMS error have an equivalent magnitude. The “percent correct”
corresponds to the percentage of patterns that have average RMS errors less than the value of
error margin.
The training and testing results are also shown in Figures 6-13 and 6-14 respectively. As seen,
the results are remarkably good.
6-8
9807070
Laboratory Test
Table 6-2
Localization Results for Running Rotor Using Neural Network
Training Statistics Testing Statistics
Average RMS Error 0.320079 0.28952
Max RMS Error 1.53572 0.961365
Max Error 1.53572 -0.961365
Number Correct 148 54
Percent Correct 46.98% 40%
Output,Target Vs. Pattern

35.75
for Node #1
Train Data Targets

28.60
21.45
Output
14.30 NN Outputs
7.15
0.00
0.00 1 37 73 109 145 181 217 253 289 315
Pattern #
Figure 6-13
The Training Result
6-9
9807070
Laboratory Test
Output,Target Vs. Pattern

35.75
for Node #1
Test Data Targets

28.60
21.45
Output
14.30
NN Outputs
7.15
0.00
1 17 33 49 65 81 97 113 129 135
Pattern #
Figure 6-14
The Testing Result
6-10
9807070
7
CONCLUSIONS AND FUTURE WORK
A system has been developed to detect and localize a short in the rotor windings of a
turboalternator. The system, based on traveling waves, provides excellent results. Hardware
circuits have been developed to generate the necessary signals. The circuits have the advantage
of being able to float on any applicable DC or AC voltage so that testing can be performed on-
line for excited windings. The system does not require installation of any hardware inside the
machine. The detection device need only to be connected to the slip rings of the synchronous
machine.
The designed hardware is high repeatable. The system is capable of handling variations in the
signature signals due to several factors such as
• machine temperature,
• speed,
• excitation current, and
• brush noise.
Rigorous testing of the technique was performed on a test setup constructed in the CIA lab of the
University of Washington. During the development process, several aspects of the system were
tested in the field. All test results show the effectiveness of the developed methods to detect and
localize the shorts in the rotor circuit.
To further refine the system, future works in this area should include the following:
1. Detection of intermittent shorts. Most of the shorts in the rotor are due to copper dust, and
initially are intermittent and speed dependant. If the detection is possible for these
intermittent shorts, preventive actions could be done to avoid the more serious permanent
shorts.
2. Detection of multiple shorts. The developed techniques can not localize multiple shorts.
To do this, a sequential process must be performed where the first short is cleared and the
system is re-tested. It is possible that the system can be modified to recognize multiple
shorts.
3. Establishment of signature signals data base. It is possible to create a database for several
classes of generators. This database can be obtained for the new machines before they are
commissioned. The results can be used as the base case for short turn detection.
7-1
9807070
Conclusions and Future Work
4. Performance of corroborative field-tests. An elaborate field-test is needed to verify the

effectiveness of the developed system. The short detection should be tested for operational
and excited machines. For localization, the test should be done when the machine is out of
service for maintenance.
7-2
9807070
8
BILIOGRAPHY
[1] D. R. Albright, “Interturn short-circuit detection for turbine generator rotor winding,”
IEEE Transactions on Power Apparatus and Systems, vol. PAS-90, 1971, pp. 478-483.
[2] J. Muhlhaus, D. M. Ward and I. Lodge, “The detection of shorted turns in alternator rotor
windings by measurement of circulating stator currents,” Second International
Conference on Electric Machines – Design and Applications (Conf. Publ. No. 254),
London, UK, pp. 100-3, 17-19 Sept. 1985.
[3] J. W. Wood and R. T. Hindmarch, “Rotor winding short detection,” IEE Proceedings,
vol. 133, Pt. B, No. 3, pp. 181-9, May 1986.
[4] A. J. Ellison and C. J. Moore, "Acoustic Noise and Vibration of Rotating Electric
Machines", IEE Proceedings, Pt. B, Vol. 115, Nov. 1968, pp. 1633-1640.
[5] L. C. Shan and J. A. Kong, Applied Electromagnetism, PWS Publishers, Boston, 1987.
[6] S. Ramo, J. R. Whinnery and T. Van Duzer, Fields and Waves in Communication
Electronics, Wiley, New York, 1965.
[7] M. A. El-Sharkawi, R. J. Marks II, S. Oh, S. J. Huang, I. Kerszenbaum and A. Rodriguez,

"Localization of Winding Shorts Using Fuzzified Neural Networks," IEEE Transactions
on Energy Conversion, Vol. 10, No. 1, pp. 140-146, March 1995.
[8] M. A. El-Sharkawi, S. J. Huang, R. J. Marks II, S. Oh, I. Kerszenbaum and A. Rodriguez,

"Neural Network Application to Short Turn Localization Using Fuzzified Data,"
International Conference on Intelligent System Application to Power Systems,
Montpellier, France, September 5-9, 1994.
[9] Robert J. Streifel, R. J. Marks II, M. A. El-Sharkawi and I. Kerszenbaum, "Detection of

Shorted Turns in the Field Winding of Turbine Generator Rotors Using Novelty
Detectors - Development and Field Test," IEEE Transactions on Energy Conversion, Vol.
11, No. 2, pp. 312-317, June 1996.
[10] Robert J. Marks II and M. A. El-Sharkawi, "Shorted Windings Sensing for Excited
Electrical Machines," IEEE International Symposium on Diagnostics for Electrical
Machines, Power Electronics and Drives, Carry-le-Rouet, France, Sept. 1-3, 1997.
8-1
9807070
Biliography
[11] S. E. Guttormsson, R. J. Marks II, M. A. El-Sharkawi and I. Kerszenbaum, "Elliptical

Novelty Grouping for On-Line Short-Turn Detection of Excited Running Rotors," IEEE
Transactions on Power Systems, in print.
[12] M. B. Zayan, M. A. El-Sharkawi and N. R. Prasad, "Comparative Study of Feature

Extraction Techniques for Neural Network Classifier," International Conference on
Intelligent System Application to Power Systems, Orlando, Florida, Jan. 28-Feb. 2, 1996.
[13] S. Weerasooriya and M. A. El-Sharkawi, "Feature Selection for Static Security

Assessment Using Neural Networks," IEEE International Symposium on Circuits and
Systems, San Diego, California, May 10-13, 1992, pp. 1693-1696.
[14] K. Fukunaga and Warren L. G. Koontz, “Application of the Karhunen-Loève Expansion

to Feature Selection and Ordering,” IEEE Transactions on Computers, Vol. C-19, No. 4,
April 1970.
[15] S. Weerasooriya and M. A. El-Sharkawi, "Use of Karhunen-Loève Expansion in Training

Neural Networks for Static Security Assessment," First International Forum on
Applications of Neural Networks to Power Systems, Seattle, July 23 - 26, 1991, pp. 59-
64.
[16] T. Kohonen, Self-Organization and Associative Memory, 2nd Ed. Berlin Heidelberg:
Spring-Verlag, 1988.
[17] J. A. Leonard, M. A. Kramer, “Diagnosing Dynamic Faults Using Modular Neural Nets,”
IEEE Expert, Vol. 8, No. 2, pp. 44-53, April 1993.
[18] Hans G. C. Tråvén, “A Neural Network Approach to Statistical Pattern Classification by

“Semiparametric” Estimation of Probability Density Functions,” IEEE Transactions on
Neural Networks, Vol. 2, No. 3, pp. 366-377, May 1991.
[19] S. Roberts, L. Tassarenko, “A Probabilistic Resource Allocating Network for Novelty

Detection,” Neural Computation, Vol. 6, No. 2, pp. 271-284, March 1994.
[20] J. M. Jolion, P. Meer, S. Bataouche, “Robust Clustering with Applications in Computer

Vision,” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 13, No.
8, pp. 791-801, August 1991.
[21] L. Parra, G. Deco, S. Miesbach, “Statistical Independence and Novelty Detection with
Information Preserving Nonlinear Maps,” Neural Computation, Vol. 8, No. 2, pp. 260-
269, Feb. 1996.
[22] Donald F. Specht, “Probabilistic Neural Networks,” Neural Networks, Vol. 3, No. 1, pp.
109-118, 1990.
[23] A. Papoulis, Probability, Random Variables, and Stochastic Processes, 3rd Ed. USA:
McGraw-Hill Inc., 1991.
8-2
9807070
Biliography
[24] K. Mehrotra, C. K. Mohan, S. Ranka, Elements of Artificial Neural Networks, USA: The
MIT Press, 1997.
[25] S. E. Guttormsson, “Novelty Detection of Shorted Turns in Turbo-Generator Rotors”,

MSEE thesis, University of Washington, 1997
8-3
9807070
9807070

Single User License Agreement

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Single User License Agreement

Transféré par

Droits d'auteur :

Formats disponibles

SINGLE USER LICENSE AGREEMENT

Final Report, November 1999

EPRI Project Manager

(A) MAKES ANY WARRANTY OR REPRESENTATION WHATSOEVER, EXPRESS OR IMPLIED, (I)

(B) ASSUMES RESPONSIBILITY FOR ANY DAMAGES OR OTHER LIABILITY WHATSOEVER

ORGANIZATION(S) THAT PREPARED THIS DOCUMENT

Computational Intelligence Applications Laboratory

This report was prepared by

Computational Intelligence Applications Laboratory

1 PROBLEM FORMATION AND SYSTEM FUNCTIONAL SPECIFICATION ........................ 1-1

2 FULLY FUNCTIONAL SYSTEM DESIGN ........................................................................... 2-1

3 SIGNATURE SIGNALS AND OUTLIER DETECTION......................................................... 3-1

4 NOVELTY FILTER FOR SHORT DETECTION ................................................................... 4-1

5 NEURAL NETWORK FOR SHORT LOCALIZATION.......................................................... 5-1

6 LABORATORY TEST.......................................................................................................... 6-1

6.1 Short Detection............................................................................................................. 6-3

7 CONCLUSIONS AND FUTURE WORK .............................................................................. 7-1

8 BILIOGRAPHY .................................................................................................................... 8-1

Figure 2-1 Basic Setup............................................................................................................ 2-2

Table 4-1 Detection results for voltage-excited running rotor................................................ 4-10

1.1 Problem Formulation

Problem Formation and System Functional Specification

1.2 General Operational Specifications

There are several design considerations in this project:

Problem Formation and System Functional Specification

2.1 Basic Blocks

2.2 Test system

Fully Functional System Design

Fully Functional System Design

Fully Functional System Design

2.3 Hardware Design

Sharp edge signals

Fully Functional System Design

2.3.1. Signal Generator

Fully Functional System Design

Fully Functional System Design

2.3.2. Signal Detector

Fully Functional System Design

Fully Functional System Design

2.3.3 Problems and Modifications

2.4 Hardware Design and Requirements for Field test

3.1 Twin and In-Out Signature Signals

Sending signal Receiving signal

Signature Signals and Outlier Detection

3.2 Outlier Detection

To eliminate significant interference in the measurements and corresponding detection

The standard deviation is estimate by

Signature Signals and Outlier Detection

(xi − x )T (xi − x ) > 3s (3.3)

Surfaces that can be placed around the healthy data include

Novelty Filter for Short Detection

Novelty detection can be couched in the paradigm of hypothesis testing.

H0: status quo (no winding shorts)

α = Probability[ H1 is announced given H0 is true ]

and the detection probability is

β = Probability[ H1 is announced given H1 is true ]

Novelty Filter for Short Detection

4.1 Computation of Detection Threshold

Novelty Filter for Short Detection

Fault = True if x − x > T

Novelty Filter for Short Detection

Novelty Filter for Short Detection

where N is the number of healthy signatures, aj is coordinate j of the hyperellipse, and xj is