tively short cable lengths between two substations that can be isolated from the grid without interrupting power to customers.
O-line method is, per denition, restricted to one cable section, more cables can't be taken out of service without having a serious outage. However, this is not a restriction at all in on-line measurement.
The superiority of on-line to o-line measurement is related to the possibility of trend watching. PD activity changes over time and these changes yield valuable information on aging progression. O-line measurements have lim- ited capability of indicating the insulation state over time. They only provide snapshots of the insulation status during the cooling down of the cable and once every certain period of time e.g. once or a few times a year, while on-line measurements are capable to monitor the insulation condition continuously to detect upcoming defects like water ingress, mechanical damages, etc., and monitor them over time.
On-line techniques are capable of capturing the PD activities that only ap- pear for a short time, which also make them superior to o-line techniques In conclusion, on-line PD measurement is non invasive and non-destructive pre- dictive test procedure, giving a representative picture of PD activity under normal load conditions. It allows observing actual trends in PD behavior which may indicate the actual insulation condition of the cable.
3.4 PD diagnostic tool - Smart cable guard
Research invested during the last decades resulted in developing various diagnostic tools for medium-voltage power cables to be used either o-line or on-line. These PD measuring tools are capable of detecting a defect in the insulation, but not all of them are suitable to identify and locate at the same time [37]. O-line techniques which are designed for defect identication and location include one sensor which performs the measurement of both the PD signal itself and its reection at the far end. Such method may be hampered by possibly weak reection if it would be applied for on-line situations, due to the load connected to the cable being close to the cable characteristic impedance. On-line techniques can employ a similar approach based on measuring with two sensors, each at one cable end (Figure 3.6).
The SCG (smart cable guard), was developed for PD monitoring in MV cable systems. This system comprises two units, to be placed at both ends of a cable connection. Each sensor is connected to a supporting computer that can commu- nicate via Internet with the data control center. Each unit consists of an inductive sensor/injector unit (SIU) and a controller unit (CU).
Inductive detection is chosen to allow for safe installation, avoiding galvanic con- tact with energized parts. The inductive sensors employed in SCG include a
coil with magnetic core which can be clamped around the cable end. Figure 3.7 schematically shows typical components including transformer, and cables, in MV substations or RMUs. The preferred locations to install the sensor are indicated with circles. Installing the sensor at these locations results in measuring only sig- nals from the cable, while providing safe installation. Each of the shown locations has its own pros and cons which are further discussed in [25, 37]).
Detection and location of PDs by the SCG system involves measurement, syn- chronization and calibration.
Measurement: The measurement is done through the sensors clamped around the cable ends with regular intervals. Each measured record is sampled with 50 MHz, sucient to cover relevant frequencies contained by PD signals which have
Figure 3.6: PD measuring system designated for on-line detection and location
Figure 3.7: Typical RMU layout with indicated potential positions to install inductive coil[25, 37]
3.4. PD DIAGNOSTIC TOOL - SMART CABLE GUARD 35 traveled hundreds of meters or more (up to 5 MHz). A record contains one full power frequency cycle, i.e. 20 ms. In each record an injected reference pulse is included. At each cable end the record is analyzed. Arrival times and magnitudes of the signals which are identied as PD events are extracted. The parameters are communicated to a control center where the information is combined to nd the PD locations. This whole process takes about one minute after which the next record can be taken. The data collected continuously are stored in a database. Synchronization: The PD location is extracted from the dierence in time of arrival at both cable ends. Therefore, both units need to be synchronized. The syn- chronization is accomplished by pulses injected through inductive coils (a patented solution) in regular basis at one cable end and measured at both cable ends [37]. The total cable propagation time is obtained and synchronization of both units is accomplished by these reference pulses [37].
Calibration: It is also required to calibrate the measuring system. To this end, the injected signals for synchronization can be used as well. In principle, the transfer of the injected waveform to the far end is measured and a model can be constructed for the frequency response. Such a model includes parameters for the signal propagation along the cable and signal coupling at the cable ends. From this model the transfer can be estimated of a PD arising anywhere in the cable. Data interpretation: Proper interpretation is needed to actually give a meaning to PD measured along the cable system. The data can be correlated to a potential type of the defect. They also can be used to estimate the status of the defect as well as the potential risk and failure time. The next chapter describes the pattern recognition techniques employed to interpret the PD data.
Chapter 4
Partial discharge pattern recognition
Condition-based maintenance, is a key issue for utilities to keep their systems running. Awareness on system state is a prerequisite for sound decisions to utilize appropriate maintenance/replacement strategies. Basically, condition monitoring tools are used to assist the network owners in avoiding unplanned outages in the electric power network. A key function of such tool is to capture signs of degradation and predict whether possible failure is on hand. Further, it is used to recommend on possible remedial actions to be taken to prevent a breakdown in the system. Condition monitoring comprises three main steps namely data acquisition, data handling and data utilization. In general, the main intention behind a data handling module is to detect abnormalities in the system, associate them with characteristics obtained e.g. from past experiences or from a training dataset and extract the underlying trend. PD events are indications of an ongoing degradation process in the insulation [4]. For an installed MV cable system, partial discharge (PD) measurement is one of the most helpful methods to evaluate the state of the insulation especially when it provides the possibility of pinpointing the PD source. The electromagnetic signals originating from the PD are capable to reach the cable ends where they can be detected [4]. Therefore, information provided through PD diagnostics is invaluable for maintenance activities. However, such raw measure is useful only if it is accompanied by an interpretation process. Thus, PDs detected continuously need to be analyzed to extract abstract measures to possibly identify the underlying degradation process as well as the existing status and ideally the remaining life of the insulation media. This chapter explores various steps to identify the insulation system's state based on PD measurement and deduce information from raw PD data.4.1 Data mining and pattern recognition
Implementing the condition based maintenance approach results in a huge data stream which need to be translated to meaningful information (knowledge). How- ever, the extraction of the knowledge is a delicate task [38]. Various techniques
can be employed to map the low level data to useful information [39]. For exam- ple, a proper model can be developed describing the degradation process. Usually, the physical degradation mechanism is complex and only partly understood. Rel- evant parameters to describe the nature behind the PD phenomena can hardly be attained. Alternatively, models can be considered that describe the dataset in an abstract way. At the core of this process lay data mining methods for pat- tern recognition and extraction [39]. Traditionally, such a process relied on human power and was done manually. However, thanks to the advances taken place in the eld of electronic computation, this can be done much faster and more accurate nowadays.
For many years, defect identication based on PD recognition was done by observing the PD pulses shown on an ellipse representing the power frequency signal on an oscilloscope screen. The advancement in electronic computation and digital monitoring [40], as in other elds, opened up new opportunities for ex- ploiting automated pattern recognition techniques to eciently manipulate the vast stream of data to diagnose potential deciencies in insulating media. These methods dier from each other rstly in the way of representing the data and secondly in the approach of classifying the defects. To properly represent the data, various approaches such as pulse shape [41, 42, 43, 44, 45, 46, 47, 48, 49, 50] presentation, statistical parameters [40, 51, 52, 53, 54, 55] and statistical distribu- tions [38, 39, 51, 56, 57] are employed. The classication of the defects are done through dierent techniques for instance by means of statistical based algorithms [40, 42, 51, 52, 58] or by means of neural networks [59, 60, 61, 62, 63, 64, 65, 66], which are nowadays common pattern recognition techniques. Generally, for the classication tool, only limited trained PD datasets developed based on a few sim- ple degradation processes from samples in laboratory environments, are available. However, for systems operating under real operational condition the aging mech- anism, the aging scale and the failure process are more complex as compared to those under laboratory conditions. Therefore, the developed trained dataset may fail to respond to the real aging process. In case of on-line PD measurement, where a variety of defects with mostly unknown background exist, another approach is being used which is described in the following chapters.