Condition monitoring of power transformer on-load tap-changers. Part 2: Detection of ageing from vibration signatures. P.Kang and D.Birtwhistle. Abstract: The ...
Condition monitoring of power transformer on-load tap-changers. Part 2: Detection of ageing from vibration signatures P.Kang and D.Birtwhistle
Abstract: The paper describes a technique for on-line automatic condition assessment of an on-load tap-changer (OLTC) using a self-organising map (SOM). With a condition indicator giving the correct indication of the current condition status, an estimate can be made of the remaining life of the equipment. The condition assessment technique is demonstrated using the signatures collected by online monitoring systems installed on selector type OLTCs in distribution substations. Using the realtime fault detection procedure, reliable identification of incipient faults in the equipment can be achieved for the pre-specified false alarming rate.
1
5
Introduction
In Part 1 of this paper series [l] we discussed the automatic condition diagnosis of on-load tap-changers (OLTCs). In this paper we address the issue of automatic incipient fault detection in OLTCs utilising Kohonen’s self-organising map (SOM) [2]. The SOM is used to handle the statistical variability of signatures of normal condition. The minimum quantisation error (MQE) of the SOM is formulated as a single indicator to give a quantified description of the changes of the equipment condition. We also develop a simplistic on-line detection scheme, which is effective in identifying both abrupt and progressive degradation. In previous work, two different approaches have been proposed for the detection of abnormalities in the vibration signatures of OLTCs. The technique of Bengtsson et al. [3] uses the average of a number of envelopes of vibration signals of a healthy OLTC as a reference, and identifies the abnormality in the vibration signatures from the difference between the reference and acquired envelope signature. This technique requires accurate temporal alignment of envelope signatures. Wright and Bushby [4] developed an alternative signature analysis technique, which uses event timing and shape analysis to detect faults in a diverterswitch type OLTC, but there is little published information regarding the field application of this technique. Previous condition assessment techniques [3, 41 have implicitly assumed that the change of OLTC condition is one abrupt step process, i.e. from normal to faulty. However, in practice, the equipment may fail in a number of ways. Faults such as cracked components may cause the condition of the equipment to deteriorate unpredictably. Failure modes such as wear of contacts may occur in an 0IEE, 2001 ZEE Proceedings online no. 20010388 DO1 lO.l049/ip-gtd:20010388 Paper frst received 17th May 2000 and in revised form 18th January 2001 P. Kang is with United TechnologieS Research Center, 41 1 Silver Lane, Fast Hartford, C T 06108, USA D. Birtwhistle is with Research Concentration in Electrical Energy, Queensland University of Technology, 2 George Street, GPO Box 2434, Brisbane, Q 4001, Australia IEE Proc.-Gener. Transm. Distrih., Vol. 148, No. 4, July 2001
.
irregular fashion. The condition of the equipment may gradually degrade for a period of time, then stay constant for a period of time, then deteriorate to a new constant level, and so on until it fails to perform its required function. In the present paper, fault detection schemes are introduced that facilitate automatic monitoring and detection of changes in OLTC conditions under those failure modes described above. 2
Fault detection scheme
2. I Condition indication As has been discussed [l], the vibration signal produced by contact movements is processed using the wavelet envelope smoothing technique. The feature data used to train a SOM is the auto-correlation function of the smoothed envelope. For fault detection application, a SOM needs to be trained using the feature data prepared from a number of signatures of normal condition. These signatures should be representative of possible variations existing in the signatures for different transformer load conditions. When newly acquired feature data are projected onto the trained map, the resultant MQE is used as a single index showing the difference between the signature and the best matching normal signature stored in the map. Higher values of MQE indicate that faults might have occurred in the equipment. Since MQE is log-normally distributed [I], the logarithm of MQE, x = ln(MQE), which is normally distributed, is used in the fault detection algorithms described in the following section. 2.2 Fault detection algorithms 2.2.1 Detection procedures: Traditional detection algorithms [5] have been formulated to detect one-step changes in mean values. However, the change of OLTC condition most often follows a gradual process rather than a single sudden jump. One way in which the measured value of x changes with time is illustrated in Fig. 1. The mean of x undergoes a monotonic process: a series of upward shifts, ,ul, ... h-l,before it approaches the critical value, pc, after which faults may start to develop in the equipment. When the mean is below the equipment is 307
regarded as being in the normal condition. As soon as p, is exceeded, an alarm should be raised with minimum detection delay, zD. equipment in faulty condition
I
0
Nl
Nn
b.1
number of operations
Fig. 9 ~
Illustrutwn of OLTC condition degruu'ution mean path
....fr... raw data
The monotonic mean path of the raw data samples is estimated using restricted regression, in which the distance of raw data samples from the mean path is minimised subject to the monotonic constraint:
not enter the detection scheme, but two consecutive outliers in a row are considered to be caused by the faults, and an alarm will be raised.
2.2.2 Threshold selection: The thresholds for the two detection procedures are selected to ensure that the rate of false alarming will not exceed a desired value. The false alarming rate is dependent on the path taken by x during equipment condition deterioration. The mean path may, however, be different even for a similar OLTC as there is a statistical variability in both the lengths of the horizontal and vertical sections of the mean path. To select the values of yc and yy, it is necessary to estimate the range of possible mean paths of x. The method proposed to facilitate the estimation is to obtain a statistical model for the mean path during the deterioration of a typical OLTC. Using the statistical model a range of mean paths may be simulated and the values of alarming threshold can be estimated for different values of false alarming probability, Pf 3
Application to a distribution class OLTC
3.1 Pre- and post-maintenance data A prototype monitoring system was installed on a selector type OLTC connected to a 60 MVA, 33/11 kV transformer
a=1 3=1
subject to p1 5 p2 5 5 pn (1) where n is the number of subgroups into which the original data series is divided, mi is the number of samples in the subgroup i, and pl, M ... ~ ( nare the corresponding mean values of the n subgroups. A numerical algorithm known as 'pool-adjacent-violators' (PAVA) [6, 71 is available for finding the maximum likelihood estimate of the mean path. This algorithm works in an iterative way by arranging the data sequence into a number of subgroups until the monotonic mean change is no longer violated. For a single jump in the mean from one constant level to another, the traditional cumulative sum procedure (CUSUM) [5] gives minimum delay in detection for a specified false alarming rate. Chang and Fricker [7] have recently modified the procedure for monotonic processes. The detection statistic of the modified CUSUM detection procedure is defined as a
.
Fig.2 Faulty condition o f f l w d contacts from tup 6 to 13 of selector type OLTC
+
c ( i ) = max{O,z(i) - pc c(i - 1)) (2) This procedure is effective in detecting small shifts of the mean above the critical mean value, pc As soon as the mean of x exceeds pc the detection statistic c(i) quickly accumulates to a larger value. Once the value of c(i) exceeds the CUSUM alarming threshold, yc, the procedure stops and an alarm is given. It is also possible that sudden large changes in the values of x may sometimes occur, which are caused by unpredictable faults such as cracked components. The detection statistic should give an immediate indication. The CUSUM procedure described above has, therefore, been extended to include the simple Schewhart procedure [SI for detection of large abrupt changes in x. The Schewhart procedure uses the instantaneous value of x as the detection statistic. This procedure stops and raises an alarm, when x(i) 2 ys, where y9 is the Shewhart alarming threshold given by y5 = pc + ka, where k is a tuning parameter and o is the standard deviation (STD) of x. The Shwehart procedure is also used to detect and reject outliers in the data samples, as outliers tend to increase the false alarming rate. As the probability of the occurrence of outlier data is normally low, a simple two-in-a-row rule [SI can be used to reduce the effect of outliers on the false alarming. This means that a single sample above ys does 308
Fig.3 Faulty condition of moving contucts of selector type OLTC IEE Pioc.-Gener. Transm. Distrib.. Vol. 148, No. 4, July 2001
for three months prior to maintenance. When maintenance was completed, the monitoring system was re-started to continuously record signatures. When the OLTC tank was opened, the fixed contacts of tap 6 to 13 were found to be worn, and the associated moving contacts were worn and pitted due to excessive arcing. The condition of the contacts is shown in Figs. 2 and 3. The vibration signatures before and after maintenance are compared in Fig. 4. It can be seen that the timing between the bursts in the faulty signature is considerably different from that of the normal signature, especially the time between the last two bursts. Since the failure mode of this type of OLTC is predominantly wear of contacts, the mean value of x before the maintenance is used as the basis to determine the critical mean value for this type of OLTC. The data collected after the maintenance were used to train a SOM, and the data collected in three months before the maintenance were applied to the trained SOM. The change of x values before and after maintenance is shown in Fig. 5.
estimated using the PAVA algorithm [6, 71. It was found that the heights of jumps in the mean path are exponentially distributed: (3)
where Ap represents the jump height and q is the mean value of the exponential distribution function, which was found to be 0.026 for the OLTC investigated. The jumps in the mean path were found to have a probability of occurrence, Pr, of: I T
Pr = 9= 0.0374 (4) Ns where Nj is the total number of mean jumps in a monitoring cycle and is the total number of sampling periods in the monitoring period. The jump periods in the monitoring cycle are distributed geometrically. The STD of the data series, o was estimated from the difference between individual data points and the estimate of mean path, and found to be 0.204.
$
0
20
0
40
60
80
1
2
3
4
5
Yc
3.89 66
3.05 55
2.94 53
2.82 51
2.70 45
TD
Table 2: Effect of detection threshold (yJ on false alarm rate ( f f and ) detection delay (zD)in number of operations for moddied Shewart procedure Pf, %
1
2
3
4
5
k
3.52
ys= p c + /CO
1.92 206
3.38 1.89
3.21 1.86
3.14 1.84
3.09 1.82
168
154
141
132
Ti7
120 140 160 180 200
a
I
Table 1: Effect of detection threshold (yJ on false alarm rate ( f f )and detection delay (zD) in number of operations for modified CUSUM procedure Pf, %
100
0
20
60
40
80
100
120 140
160 180
200
time, ms
b
Fig.7 Si&atures of tap operationfrom 7 to 8 offaulty andnormal conditiom
a Faulty condition 0 Normal condition
2.4-
2.2-
#--
2.0-
8 1.8I
8 1.6-
Occasionally bad signatures caused by mis-trigger of the data acquisition system lead to outlier data in x. The probability of occurrence of bad signatures is -0.02% in the present application, which was estimated from the historical data. All outliers have been successfully rejected by comparing the instantaneous x value with the value of the Shewhart threshold. The Shewhart procedure uses the instantaneous value of x to give a fast detection of large mean shift above the critical value. Although the CUSUM procedure optimally detects small shifts of mean above the critical value, it requires the statistical model to select the threshold. When data for the model are not available, we rely on the Shewhart procedure to detect the change of equipment condition until statistical data are obtained by on-line monitoring. In this case, the alarming threshold for this procedure is selected to be one standard deviation above the critical mean value: yy = pc + cr.
3.3 Field applications 3.3.1 Abruptly occuring faults: A tap-changer monitoring system has been installed on another 33111kV transformer since mid-1999. The OLTC was maintained on 14 September 1999. In the beginning of October 2000, the monitoring system gave warning that a fault had suddenly occurred in the equipment. The abnormal signature associated with the warning is shown in Fig. 7a, in which one of the bursts normally appears in the normal signature (Fig. 7b) is missing. This indicated that a set of contacts was not making or contacting properly, causing improper contact positioning, which could be either caused by a missing contact or by problems in the driving mechanism. As this faulty signature appears only on occasions between signatures of normal condition, the missing contact hypothesis was excluded. A possible cause could be the slipping of 310
Q)
i! 1.4-
9
5
1.2-
E 1.0-
0.41 1
i
"
401
"
b
L
G
1201 1601 number of operations
801
2001
2401
Fig.8 Change in mem vulue of condition indicator after 2400 tap-change operations
In a subsequent inspection, after the wheel shaft was disassembled, the slippage in the shaft has been identified in the keyway of the Geneva wheel. There was about 1-2mm of play within the keyway, which, when taken at the extremities of the moving contacts, equated to approximately 40" of play. Considering that the distance between the fixed contacts is -5Omm, if this condition had been left unchecked, then a catastrophic failure would have been inevitable.
3.3.2 Delay of maintenance: An OLTC of the same type has been scheduled for maintenance three times in the past, and no degradation was observed at each of the maintenance inspections. In the most recent inspection, the OLTC was monitored before the scheduled maintenance, and the condition indicator of the monitoring system was found well below the alarming threshold. This indicates that the equipment was still in a healthy state. In this case all three field maintenance inspections could have been deferred by application of the monitoring system. In a similar case a monitoring system has been monitoring the condition of a tap-changer since the previous maintenance. The change of mean values of the condition indicator, MQE, for tap operation from tap 6 to 7 over IEE Proc-Gener. Transm. Distrib., Vol. 148, No. 4, July 2001
2400 operations is given in Fig. 8. Since the critical mean value of MQE for this type of OLTC is 3.32, this OLTC is still in the healthy state. The on-line monitoring of this OLTC has enabled the maintenance to be deferred by 1.5 years at this time. 4
1
{
Discussion
Although the on-line fault detection technique was developed based on the data from a type of distribution class OLTC, it is applicable to other types of equipment including transmission level OLTCs. When this detection technique is extended to other types of OLTC it will be necessary to invest a period of time in monitoring vibration signatures and relating them to the degradation of the equipment so that the critical mean value and the alarm threshold can be determined for the equipment. Older type OLTCs like the one examined in this paper are causing significant maintenance expenditure for utilities. The maintenance interval has been reduced to two years and even so there have been a number of catastrophic failures. The potential usefulness of the methodology can be seen with reference to Fig. 6b, which shows the range of possible paths that x may take. With a critical mean value of 1.2 it can be seen that this value may be exceeded at any point between 1550 and 2250 operations depending on the path of the mean. By tracking the path it is possible to defer maintenance of OLTC until a time just before it becomes essential. Under a periodic maintenance regime, maintenance may not be done before failures occur for some conditions or maintenance may be conducted too early. It may be not economical to install one monitoring system on each tap-changer, especially in distribution substations. The approach taken here has been to use the monitoring system for a period of time before and after maintenance to check that maintenance is needed and to ensure that maintenance has resulted in return of equipment to the normal condition. To obtain a useful database of normal and degradation signatures it has, however, been necessary to undertake long term monitoring of some OLTCs. Substantial savings have already been made using prototype monitors, and greater savings will be possible using the commercial units. Further research into the economics of OLTC monitoring is currently being conducted by the authors and will be the subject of future publication. 5
Conclusions
Through the utilisation of the SOM we have formulated a single condition index, the minimum quantisation error, to indicate the condition of a selector type OLTC. We have identified that the condition change under the failure mode
IEE Proc.-Gener. T i ”
Distrib., Vol. 148, No. 4, .luly 2001
of contact wear is normally a monotonic degradation process. We have developed an on-line detection algorithm, which utilises the advantages of the CUSUM and Shewhart detection procedures such that both the small shifts in the mean caused by the incipient faults as well as the large changes caused by the serious faults can be identified. We have used the on-line monitoring system to acquire data continuously from the field over several years and correlated them with observations regarding the actual condition of the equipment; the data have been obtained from the field staff following normal maintenance intervals. Using these data we are now in a position to confidently set the value for the critical mean value of the condition indicator and the probability of false alarm for the type of equipment on which our investigation has focused. As there are many OLTCs of this type, the present monitoring system is being used in a number of locations. 6
Acknowledgment
Financial support from the Queensland Electricity Transmission and Distribution Chair in Electricity Asset Management is gratefully acknowledged. The authors would also like to express their appreciation to Messrs David McCulloch, John Daly and Trevor Hubner of Substation Standards Department of Energex for considerable assistance with measurements and for providing insights into failure modes of the equipment. The authors would like to express their sincere appreciation to the research team of the Laboratory of Computer and Information Science of Helsinki University of Technology for providing access to their SOM Matlab Toolbox. References KANG, P., and BIRTWHISTLE, D.: ‘Condition monitoring of power transformer on-load tap-changers: part 1 automatic condition diagnostics’, IEE Proc., Gener. Transm. Distrib., 2001, 148, KOHONEN, T.: ‘The self-organising map’, Proc. IEEE, 1990, 78, (9), pp. 14641479 BENGTSSON, T., KOLS, H., MARTINSSON, L., FOATA, M., LEONARD, ,F., RAJOTTE, C., and AUBIN, J.: ‘Tap changer acoustic monitoring’. Presented at 10th Inteniatioiial Symposium on High voltage engineering, ISH’97, Montreal, Canada, August 1997 WRIGHT, S.E., and BUSHBY, J.P.: ‘Time domain diagnostic techniques, applied to a transformer load tap-changer’. Presented at 32nd University Power Engineering Confec$nce, UPEC‘93, Manchester, UK, 1997, pp. 253-155 MONTGOMERY, D.C.: ‘Introduction to statistical quality control’ (Wiley, 1996) ROBERTSON, T., WRIGHT, F.T., and DYKSTR, R.L.: ‘Order restricted statistical inference’ (Wiley, 1988) CHANG, J.T., and FRICKER, R.D.: ‘Detecting when a monotonically increasing mean has crossed a threshold’, J. Quu1. Techno/., 1999, 31, (2), pp. 217-234 LUCAS,,J.M., and SACCUCCI, M.S.: ‘Exponentially weighted average moving control schemes: properties and enhancements’, Technometrics, 1990, 32, (I), pp. 1-12
311
