2000 Conference on Electrical Insulation and Dielectric Phenomena

Artificial Intelligence Methodology for Separation and Classification

of Partial Discharge Signals

A. Contin A.Cavallini, G. C. Montanari and G.Pasini F. Puletti

University of Trieste LIMAT-DIE, University of Bologna TechImp
Via A.Valerio, 10 Viale Risorgimento, 2 Viale Risorgimento, 2
Trieste IT 34127 Bologna IT 40136 Bologna IT 40136

ABSTRACT Using these techniques, the fe,atures of different PD

phenomena simultaneously actike can be mixed in the
Results of investigations performed in order to improve different histograms. Complex procedures of separation
the current diagnostic techniques used for the evaluation and identification can be then necessary to extract these
of insulation systems of HV apparatus are presented in features during the data-processing stage, after PD
this paper. Improvements come from the development of measurements. Moreover, the information provided by
a new measuring system which allows the digital the rise-time and frequency content of the PD-pulse
acquisition of Partial Discharge (PD) signals and a new signals, which may help considerably in the separation,
separation method, based on a Fuzzy Classifier, for the identification and, eventually, location of the PD source,
analysis of the PD-pulse shape signals. The is lost.
identification of the classes, relevant to different PD
phenomena, is then performed by means of PD-pulse In order to improve the current diagnostic techniques, a
height and phase analysis. The proposed approach is new measuring system which allows the digital
supported by the analysis of PD data obtained from acquisition of PD-pulse signals at a sampling rate high
insulation systems of stator bars with artificially- enough to avoid frequency aliasing and with an amount
reproduced defects. of PD-pulses which enables PD stochastic analysis, has
been developed and is presented here. Separation and
INTRODUCTION classification of different PD signals are performed
Diagnosis of insulation systems of HV apparatus, by resorting to the analysis of the PD-pulse shapes by
means of Partial Discharge (PD) measurements, is means of a Fuzzy Classifier (FC), as well as to a
generally performed to infer defects and their correlation stochastic model based on the Weibull function applied
with service reliability. Most of the methods adopted for to the PD Height Distribution (PDHD). The proposed
the analysis of PD features assume implicitly the approach is supported by the analysis of PD data
presence of only one PD phenomenon (i.e., only one obtained from sections of mica-based stator bars where
source of PD or different sources of the same nature). In typical defects occurring during normal operations were
operating conditions, however, more than one source of simulated.
PD can be active due to the presence of either a plurality PD SIGNAL CLASSIFICATION TECHNIQUE
of defects or electrical noise or both. Moreover, many
insulation systems are designed to withstand a moderate PD Pulse Detection
level of PD occurring in distributed micro-voids The measuring system for the digital acquisition of PD-
(internal discharges) of the impregnatiodencapsulation pulse signals was developed resorting to a waveform
resin. Since the features relevant to the different PD analyzer which provides large bandwidth (1 GHz) and
phenomena simultaneously active can be combined in a fast sampling rate (up to 4 GSds), with triggerable
mixed PD-pulse sequence, an efficient separation partitioned on-line memory length (8 MB). This
technique is required in order to avoid misleading oscilloscope presents a sequence acquisition mode that
evaluations during the identification of the defects allows multiple events to be stored, each of them
generating PD phenomena. recorded in segmented acquis,ition memories. This
The commonly used instrumentation is based on the solution, associated with appropriate acquisition
digital PD-pulse peak detection [l]. This type of techniques, [6], permits matching of the divergent
instrumentation allows a collection of a large number of requirements of high sampling rate and large amount of
PD-signal amplitude values and provides the PD-pulses (enough to process statistically PD amplitude
measurement results in a 3D plots, that is number of and phase) to be obtained.
pulses as a function of their amplitude and phase [2].
The stochastic analysis of PD-pulse height and phase PD Pulse Classification and Clustering
distributions is currently used for diagnosis purposes [3- The time behaviour and the frequency content of each
51. PD pulse are analyzed using a Fuzzy Classifier (FC).

0-7803-6413-9/00/$10.002000 IEEE 522

 The FC can be conceptually splitted into two functions: The feature classifier can be defined as a set consisting
(a) PD-pulse feature extractor and (b) feature classifier. of a parameter vector, 8, and by a rule. Generally (but
not necessarily), 8 represents points in the data space:
The feature extractor function consist of a mapping the class prototypes or centroids, zj, that is 8={zj}. For a
which provides a compact and meaningful conventional (crisp) classifier, the rule allows to decide
representation of the measured pulses. Several mapping if a pattern, xk, belongs to class i or to some alternative
methods are available. Two of them have been class. Generally the rule resorts to the distance induced
considered here, that is, the Principal Component by some norm (euclidean, infinity, diagonal,
Analysis (PCA) and a method based on conventional Mahalonobis, city block, etc.) in the data space: given 8,
pulse-signal parameters. if zi is the prototype closest to xk, then x k is said to
In PCA, the subset of eigenvectors of the signal belong to class i, otherwise xk will belong to some other
covariance matrix provides a basis for compact pulse class.
representation [7]. In practice, each pulse is projected A fuzzy classifier does not provide such a strict partition
into the signal subspace, that is, the spanning set defined of the data space: on the basis of the distance between
by the chosen eigenvectors. However, outliers can patterns and prototypes, the algorithm provides the
negatively affect the classification. In particular, noise degree (or membership), p ~ that, xk belongs to the ith
pulses occurring at a very low rate can significantly class. Indeed, the same pattern belongs to all classes, but
affect the covariance matrix. Moreover, noise pulses, with different degrees. The sum of all membership is
whose energy can be spread over the entire space, can forced to be equal to 1:
be mapped into the signal subspace as a low energy
pulse and, thus, be misclassified [8]. Finally, PCA is
computationally intensive so that long processing times
j= 1
are required to deal with pulses acquired at a sufficiently
high sampling rate. These aspects led us to discard PCA The classification scheme here proposed is based on
for feature extraction. fuzzy algorithms. In particular, a two-stage procedure
based on the Fuzzy K-Means (FKM) and Gustafson-
The other method here envisaged is based on a rather Kessel (GK) clustering algorithms has been devised
simple mapping scheme, which uses the following
quantities relevant to time and frequency features, to, T [lo].
and W [9]: The use of two algorithms for classification purposes
- needs to be explained. In selecting a clustering
rx (r >” d t I(t - t,)Zx(t)2dt algorithm one must bear in mind that the norm used for
evaluating the distance between pattern and prototypes
to = - _ T2 = --
strongly affects the cluster shapes. In particular, k-means
(x(t)Zdr jx(t)’dt and FKM algorithms generate spherical clusters. Since
- _ -_ the clusters of pulse features here discussed seldomly
resemble spheres, then the use of a more complex
clustering algorithm becomes necessary. In particular, as
a trade-off between flexibility, robustness and
computation speed, it was chosen to use the GK
algorithm which, resorting to the Mahalonobis norm,
generate elliptic clusters in the data space. Other
alternative schemes, such as the fuzzy maximum
where uppercase indicates Fourier transform.
likelihood were also tested, but the results were not
These quantities have interesting properties: is the encouraging due to numerical instabilities.
time position of the pulse, T is the equivalent length and
However, in all clustering algorithms, the parameter 8
W the bandwidth. Their computation is quite fast and is
needs be initialized to a guess value eo. Iterative
not affected by outliers since no statistical processing is
procedures permit parameter tuning, so that the classifier
here required. One main disadvantage is that a point
parameters (centroids) can be optimized in some sense.
(b,T,W) is not directly associated with any specific
Indeed, the GK algorithm usually display a poor
waveshape. A link must be obtained by “a posteriori“
performance if not properly initialized (i.e., 8 0 should be
considerations on the available database. Also, the link
as close as possible to 8, so that a robust initialization is
between x(t) and X(f) (via the Fourier transform)
strongly advised). On the contrary, FKM algorithms
provides the correlation between T and W. In summary,
display quick convergence rate, almost independently
although this mapping scheme is rather simple and
from the guess set 80, which can, therefore, be selected
appealing, significant improvements are needed.

523
randomly in the data-space. The independence between
eo and the tuned parameter vector makes FKM
algorithm the technique mostly employed for obtaining a
reasonable guess for more complex algorithms as the
GK one here employed. These considerations on cluster
shape and initialization explain why the FKM algorithm
has been here employed to provide an initial guess for
the parameter vector relevant to the GK algorithm.
The proposed procedure includes a technique for
automatically choosing the number C of prototypes.
This technique is based on the fact that, if a vector is
poorly represented in all classes at the same time, then, 00_10_I_MER_AIR_LR_01 a
in the worst case, all membership values will be equal to
1/C (that is, the minimum value which can be achieved
simultaneously by all membership values, whose sum is
bounded to be equal to 1) and the variance (i.e., the sum
of the squares of the differences between the
memberships and the average membership) of the
memberships, sk, will be zero. The maximum variance is
attained when the pattern is coincident with a protot pe.
In this case it is easy to show that sk=(l-l/c) A
criterion for deciding whether or not a pattern is badly
4
represented in all classes can thus be based on a b C

minimum value for sk: if min(sk) is < ~ ( I - I / c ) * ,being

O<q<l, then the classifier performs poorly on xk. If
parameter q is equal to 1, a class is created for every
vector xk. On the contrary, if q is close to 0, only a few
classes are created. For practical purposes, q was set
equal to 0.7, according to empirical observations on
classifier performances. The result of Fuzzy
Classification stage is a number of classes that include
PD signals having homogeneous shape.
Evaluation of classification effectiveness and PD- e
d
process identification
Fig. 1 3D PD patterns due to internal discharges and
The evaluation of the sub-set of PD-pulses belonging to electical noise before (a) and after (ti), (c) FC application. PD
each class is performed in this paper, for the sake of signals relevant to internal PD (d) and electrical noise (e),
brevity, resorting to stochastic PD-pulse height analysis detected on stator bar #1 at 7 kV.
(however, also the phase analysis and the polarity
phenomena are generally considered). The classification occurring during normal operations were simulated. In
effectiveness of the FC can be checked assuming that particular, PD phenomena due to: (A) internal
different PD phenomena can produce different PDHD. discharges and electrical noise, (:B) surface discharges,
Under this assumption, the whole set of experimental (C) discharges due to delamination and (D) multiple
data, as well as each cluster, can be processed according surface PD discharges, were artificially reproduced in
to the two parameter Weibull function [ 3 ] .The values of order to test the procedure in well-known conditions.
the scale, a,and shape, p, parameters can be used for
PD-source indentification, [ 3 , 61, while the linearity The data examined here was collected from three
(Cramer-von Mises, CVM)test can assess if each class different stator bars (named #1 to #3 which correspond
collects data representing a single PD phenomenon, i.e. to defects A, B, C, respectively) and a coil (#4, defect
discharge pulses coming from a given typology of PD (D)). The bars have the same length (1.5 m) while the
source. coil size is (1.OxO.4 m). The insulation system is based
on mica-tape impregnated with epoxy resin but with
EXPERIMENTAL APPLICATIONS different insulation thickness (with design electrical
stress close to 3 kV/mm) and different number of
The proposed approach was applied to analyze PD data conductors.
detected on AC stator bars where typical defects

524
PD measurements were performed at different voltage
levels starting from the nominal stress, i.e. from 4.5 to
11 kV. Due to the influence of the longitudinal
inductance on signal shape, calibration was not
performed and PD results are provided in Volt, instead
of picoCoulomb. For the sake of brevity and in order to
highlight the FC method, the results here reported are
relevant to a single level of applied voltage.
Stator bar #1 was apparently devoid of macro-defects.
PD were due, indeed, only to micro-voids included in
the insulation system. PD measurements were performed
adopting a low trigger level in order to collect also the 00_10_I_MER_AIR_LR_02 a 00_10_S_MER_AIR_LR_01
pulsanting background noise. The relevant 3D PD plot,
detected at 7 kV, is reported in Fig.la. The application
of the FC method allowed a partition of the PD-pulse
signals in two classes (reported in Figs.lb and IC. The
comparison between the 3D pattern and the signal
shapes detected in the presence (Figs.lc and le) and in
the absence of the applied voltage confirms that the FC
classifies correctly the PD and the noise. The analysis of
the parameter values of the Weibull model shows that
the 3D plot of Fig.lb is relevant to PD due to micro
voids (the value of ranges, in fact, between 2 and 3). b C
An example of PD signal is reported in Fig. Id.
The corona suppressing system of one side of stator bar
was completely removed in order to reproduce two PD
phenomena completely overlapped in terms of PD-pulse
height distributions. This effect was obtained applying
4.5 kV; the 3D pattern is shown in Fig.2a. The FC was
able to distinguish the two PD phenomena producing
two classes (Figs.2b and 2c). The presence (Fig.2d) and
the absence (Fig.2e) of oscillations in the PD-pulse
shapes is associated with surface and internal PD for the d e
two classes, respectively.
Fig.2 3D PD pattems due to intemal and surface discharges
A surface of 1 cm2 of LV electrode of bar #3 was before (a) and after (b), (c) FC application. PD signals
removed and covered with a metallic foil in order to relevant to intemal (d) and surface (e) PD, detected on
reproduce a delamination area, located in the middle of stator bar #2 at 4.5 kV
the bar. The analysis of PD measurements, performed at (internal and surface discharges). The 3D PD pattern,
5 to 8 kV with steps of 1 kV, by means of the two- detected at 6 kV, is reported in Fig.4a. The FC method
parameter Weibull function and the CVM test, showed provided a partition in two classes: the first one shows
the inception of a new PD phenomenon at 6 kV. The FC an irregular 3D plot, (Fig.4b), while the second one
method was not able to distinguish the different PD presents a quite regular PD pattern (Fig.4~).An example
phenomena, since the PD-pulse shape were quite similar of PD pulses belonging to the first and the second class
(Fig. 3c). The comparison of the 3D plot of Fig.3a and are reported in Fig.4d and Fig.4e, respectively. The FC
Fig.lb (relevant to internal discharges only) clearly classifies correctly the multiple PD sources, due to
indicates the presence of a combined PD phenomena. In corona suppressing degradation, in a single class.
fact, the CVMtest failed and the two-parameter Weibull
function cannot fit the experimental PDHD (Fig.3b). CONCLUSIONS
Hence, classification can be carried out, in this case, by
the five-parameterWeibull distribution (see [3]). A Fuzzy Classifier algorithm, developed for separation
and classification of PD signals, has been presented in
One of the four corona-suppressing tapes of the coil #4 this paper. If separation is performed successfully, the
was short circuited using a conducting adhesive tape in original PD-pulse sequence can be splitted in more sub-
order to reproduce a defect with multiple sources
sequences, each one relevant to a well defined PD

525
 a 00_10_I_MER_AIR_LR_03 a 00_10_S_MER_AIR_LR_02

b C b C

Fig.3 3D PD pattern due to intemal and delamination

discharges (a), the Weibull graph relevant to positive PD,
and a PD signal detected on stator bar #3 at 6 kV
phenomenon or to electrical noise. The FC method can
fail when different PD phenomena produce quite similar
signal shapes. In the latter case, the stochastic analysis
of the PDHD can detect the presence of a plurality of
PD sources and, possibly, carry out separation and
identification. It is noteworthy that the proposed FC can e
d
operate efficiently as a rejection filter for external noise.
Fig.4 3D PD pattems due to internal and surface discharges
REFERENCES before (a) and after (b), (c) FC application. PD signals
relevant to internal (d) and surface (e) PD, detected on
[ 11 B.H.Ward, “Digital Techniques for PD stator bar #3 at 6 kV.
Measurements”, IEEE Trans. on Power Delivery, V.7,
N.2, pp.469-479, April 1992. [6] G.C.Montanari, A.Contin, A.Cavallini, ”Random
[2] B.Fruth, and J.Fuhr, “Partial Discharge Pattern Sampling and Data Processing for PD-Pulse Height and
Recognition - A Tool for Diagnosis and Monitoring of Shape Analysis”, IEEE Trans. on Dielectrics and
Aging”, CIGRE Report, Paper 15/33-12, Paris, France, Electrical Insulation, V.7, N.1, pp.30-39, February
1990. 2000.
[3] A.Contin, G.C.Montanari, C.Ferraro: “Partial- [7] H.Borsi, “A PD Measuring and Evaluation System
Discharge Source Recognition by Weibull Processing of Based on Digital Signal Processing”, IEEE Trans. On
the Pulse-Height Distribution”, IEEE Trans. on Dielectrics and Electrical Insulation, V.7, N. 1, pp.21-29,
Dielectrics and Electrical Insulation, vo1.7, N. 1, pp.48- February 2000.
58, February 2000. [8] B.D. Ripley, Pattern Recognition and Neural
[4] T.Okamoto, and T.Tanaka, “Novel Partial Networks, Cambridge University Press, Cambridge
Discharge Measurement Computer-Aided System”, (UK), 1996.
IEEE Trans. on Electrical Insulation, v01.21, n.6, [9] L.E.. Franks, Signal Theory, Prentice Hall,
pp.1015-1019, December 1986. Englewood Cliffs, USA, 1975.
[5] E.Gulski, Computer-Aided Recognition of PD Using [IO] R. Babuska, Fuzzy Modeling for Control, Kluwer
Statistical Tools, Ph.D Thesis, Delft University Press, Academic Publishers, Boston (USA), 1998.
1991.

526