UHF Partial Discharge Monitoring System in Service: Generating and responding to PD alarms F. I. Cook* R&D Department DMS- Glasgow, UK [email protected]

Abstract—the current paper will discuss how Artificial network, from PD signal, to PD alarm, to investigation and intelligence techniques can be used to create an Expert System rectification. that will intelligently generate alarms in the presence of Partial Discharge (PD) activity within Gas insulated (GIS). In II. PDM SYSTEM addition the appropriate practical response to those alarms for The following diagram (Fig.1) outlines the basic components in-service GIS is explained and discussed relative to a practical example. of a Partial Discharge Monitoring (PDM) system.

Keywords—Partial Discharge; GIS; UHF; Condition Monitoring; PD Diagnosis; Expert System; Artificial Intelligence; PD Alarm; Artificial Neural Networks ( ANN); genetic algorithms;

I. INTRODUCTION The most common causes of electrical failure in GIS generate Partial Discharge (PD) activity in advance of complete breakdown [2]. It has been shown that the UHF method provides a sensitive and reliable method of detecting PD [4].The UHF method developed over 20 years ago in the UK [3] has been adopted worldwide and is in common use by GIS manufacturers and utilities. The microwave resonances excited in the GIS chamber by the PD pulse (rather than the actual PD current pulse) can be Fig. 1. General arrangement of a PDM System picked up by UHF couplers installed in the GIS. Different PD sources have different and identifiable UHF pattern The UHF data can be presented such that the significant and characteristics [3]. The UHF pattern can therefore be unique characteristics of a particular defect type can be seen. interpreted (by human expert or automated expert system) to Data can be shown in a number of other useful and meaningful identify the nature of the source of the PD signal [6]. representations including: Single Cycle (PRPS); POW (Peak There are established benefits of installing a UHF PDM system Hold); 24 hour History; Phase Resolved Partial Discharge to limit the potentially serious consequences of breakdown, in (PRPD). Each different data representation provides terms of unplanned circuit disruption, penalties due to loss of complimentary information to classify and assess the PD supply and cost of repairing damage to GIS [7]. activity taking place (Fig.2 and Fig.3).

To benefit from the installation of a PDM system the data collected by the system must be acted upon appropriately. It is detrimental to ignore PD signals and take no action when action is required. It is also detrimental to take outages and open up GIS unnecessarily. The use of Artificial Intelligence to automatically classify PD signals and generate alarms will be discussed. The Expert System has a high degree of accuracy and has been installed on Fig. 2. Examples of different display types hundreds of substations worldwide. The methods and processes developed are used to provide technical advice and reporting to a number of industry customers. The operational aspects of responding to PD activity within the network will be discussed. This will be demonstrated by discussing a case study on the a sufficient database is the most critical factor when creating an expert system for non-trivial complex, classification tasks. B. Artificial Neural Network The most important classifier of the Expert System used here is the Artificial Neural Network (ANN). An ANN is a computational model based on the structure of a biological neural network e.g. the human brain (Fig.5). Fig. 3. Examples of different display types

III. EXPERT SYSTEM Different types of PD are identifiable from their signal patterns (Fig.4). If it is possible for a human expert to classify PD source type from the pattern, then it is possible to use Artificial Intelligence techniques to create an Expert System that is capable of doing the same task.

Fig. 5. Neuron in human brain and Artificial Neural Network

ANNs are excellent classifiers for use in pattern recognition tasks, where the relationship between input and output is complex. The ANN ‘learns’ by being taught a database of exemplars. The ANN will incrementally change its internal connections between its’ ‘neurons’ based on ‘experience’. This is broadly similar to the way the human brain will learn to recognise patterns. Provided the database that the ANN has learnt is sufficiently representative, the ANN will be able to correctly classify new unseen data based on its previous experience. C. Genetic Algorithm Genetic Algorithms are a computational technique based on Fig. 4. patterns from different PD signal types (Single Cycle display) biological evolutionary theory. The population starts with a random selection of possible solutions. Each solution is given The most common causes of electrical failure in GIS are: free a ‘fitness score’ based on how well they meet the metallic particle; discharges from any stress-raising protrusion characteristics of an ideal solution. The higher an individual (e.g. busbar corona); capacitive sparking from an electrode solution’s fitness, the more likely it is to ‘breed’, where as the which is not properly bonded to either the less fit functions ‘die’. The ‘population’ of different possible conductor or earth (Floating electrode) and insulation defects solutions will then ‘breed’, ‘compete’, ‘die’ and ‘mutate’; until (e.g. Void in solid ) [2]. the population converges and a solution has ‘evolved’ (Fig.6). The Expert System is able to automatically classify signals. Genetic Algorithms are excellent for optimisation tasks where The following sections discuss the major considerations for we know what characteristics an ideal solution would have, Expert System design and describe a Modular Hybrid Expert but the search space is too large to exhaustively search all System design. possible solutions. A. Database One of the most important requirements in creating an Expert Evolution: System is the database of signals. The database must be Initial selection, sufficiently large and representative of the signals that the Population population of mutation, Expert System will encounter in service. The database is used converges on potential mating, both to test and ‘train’ the Expert System. The database in this solution case consists of over 3 million exemplars, built from data from solutions crossover, live installations from different manufacturers collected over death. 15 years. A sufficiently large database is essential, not only to truly represent the scope of signals that will be encountered in Fig. 6. Basic process of genetic algorithm computation service, but also for training Artificial Neural Networks (ANNs). As the complexity of ANN required to carry out the D. Rule based classifier classification increases, the size of the training set required for Where possible the best classifier to use is often the simplest. the ANN to ‘learn’ the data increases exponentially. Arguably, If the mathematical relationship between input and output is easily described in a simple rule, then it is more efficient to use this basic method rather than the more time consuming machine learning approaches such as ANN. Raw Input E. Feature Selection Noise For non-trivial pattern recognition problems, it is not practical Removal or desirable to use raw data as the input vector to the expert system. The more inputs that are fed into the classifier, the Feature more the classifier is influenced by the infamous ‘curse of Extraction dimensionality’. This holds that as the number of inputs increases, the size of database required to train the system Other Rule Based Modular Genetic increases exponentially. Therefore, features are extracted Classifiers classifier ANNs Classifier from the raw data. Features are parameters that can describe the raw data at a higher level. Features can vary from the simple, e.g. mean pulse height; to higher level statistical parameters such as Combination Fractal Dimension. The features form the input vector to the Function classifiers. It is important that the chosen features provide enough information for the classifier to reliably map from input vector to desired classification, while at the same time Rule Based not introducing unnecessary complexity to the classifier. Executive Control Selecting appropriate, descriptive and complementary features System is extremely important when designing an accurate classifier. F. Modular Hybrid Expert System Output

The Expert System combines a number of different Artificial Fig. 7. Hybrid modular expert system Intelligence techniques (Fig.7). Different classifiers have different strengths and weaknesses. It can be difficult to find an individual classifier that provides The expert system will classify the input signal as a type of optimal results in all circumstances. It is possible to combine PD signal or a type of noise signal. multiple classifiers to create high performance classification systems. The rationale behind multiple classifier systems is, to IV. GENERATING PD ALARMS combine the different strengths of the individual classifiers to The PDM system is typically configured to generate PD make a hybrid classifier that is greater than the sum of its alarms based on the output of the Expert System. The system parts. The PDM Expert System combines various pattern will generate alarms only on those signals classified by the recognition techniques to form a highly accurate hybrid Expert System as PD. Signals classified as Non PD will not classifier contribute to alarms. When the signal activity detected by a channel has sufficient The hybrid modular expert system functions in the following pulses above pre-set amplitude then an ‘event’ will be way : triggered. The pulses contained in that event are then 1) Noise is removed from the raw input data. classified by the expert system and recorded (along with time, 2) A variety of descriptive and complimentary date, channel ID). Typically, PD alarms are initially features are extracted configured to trigger if a preset number of events per rolling 3) The extracted features form the input vector for a 24 hour window are classified as PD (this is tailored to the number of classifiers individual configuration of the station). The alarm message 4) The outputs of the classifiers are combined using generated tells us which station, sensor, time of alarm, and the an intelligent combination function. conditions that triggered the alarm (Fig.8). There is also an 5) The entire process is monitored by an executive alarm message triggered via SCADA. control system that can bypass, and modify all of the preceding stages. 6) The executive control system makes the final decision on the classification.

Fig. 8. Example of PD alarm message

V. RESPONDING TO PD ALARMS When a PD alarm is triggered, the activity that triggered the alarm is assessed based on the following criteria: • New signal or pre-existing signal • PD Signal Type • Channels active • Amplitude Fig. 9. 24 hour Peak history day data (initial) and after 6 weeks • Discharge rate • Events per day • Persistence (i.e. is signal continuous or intermittent) • activity trend • location Determining the correct course of action when a PD signal is detected is not a trivial task. It is not always necessary when PD is detected, that an outage should immediately be taken and the GIS repaired. In fact this is very rarely the response when assessing recommended action to a PD signal. Maintaining integrity of supply needs to be balanced with protecting the GIS against possible flashover when PD is present. Should a signal be detected/PD Alarm triggered, there are different levels of action in response: Fig. 10. Three different data respresentations of the PD signal. 1) Take no action- signal not PD 2) Monitor- signal not currently critical, monitor activity VII. CONCLUSION and be ready to take action should the signal start to Artificial Intelligence tools and processes have been utilised to become significant. create an accurate Expert System able to successfully classify 3) Leave section in service, monitor signal and put plan PD Signals. The Expert System is currently in use in hundreds in place to investigate and repair source at earliest of substations worldwide. The analysis and PD alarms practical time. generated by the Expert System are used to drive intelligent 4) Take immediate outage. condition monitoring of assets. It is important to take a

balanced and objective approach when reacting to the Ideally all considerations should be discussed with input from presence of PD activity within switchgear. The information OEM experts, PDM experts and the Utility. the system provides must be analysed and if necessary acted VI. CASE STUDY upon. The presence of PD in an asset does not necessarily mean that immediate and critical action needs to be taken, The following case study is from a 400KV GIS. however all PD alarms should be considered and assessed and A PD signal became active in service. When the signal was when required, appropriate action taken. This will help to first active, signal was intermittent over a 24 hour period ensure the integrity of supply and confidence in the HV (Fig.9). As time progressed the signal became less intermittent network. and more continuous (Fig.9). The signal pattern was typical of PD (Fig.10). As the signal was classified by the Expert system REFERENCES as PD and there were sufficient events, the signal activity [1] D Kopejtkova, T Molony, S Kobayashi, I M Welch. “A twenty-five became sufficient for automatic PD alarms to be triggered. year review of experience with gas-insulated substations”, CIGRE paper The PD signal’s activity was steadily increasing, and signal 23-101, Paris 1992. was becoming more and more persistent. As there was [2] B F Hampton and R J Meats. “Diagnostic measurements at UHF in gas confirmed PD activity with increasing activity, and the signal insulated substations”, Proc IEE, Vol 135, Pt C, No 2, 137-144, 1988. was consistently active, the decision was made to investigate [3] CIGRE WG 15.03. “Diagnostic methods for GIS insulating systems”, CIGRE Paper 15/23-01, Paris 1992. and correct the source of the signal. [4] J Gorablenkow, T Huecker, U Schichler, “Application of UHF Partial Relative amplitude of the signal at different sensors within the Discharge Monitoring and Expert System Diagnosis”, IEEE ISEI pp61- GIS and investigation on site by energising different sections 64, USA 1998. of GIS, suggested surge arrestor was the source of the signal. [5] N Achatz,. J Gorablenkow,. U Schichler,. B Hampton,. J Pearson. Replacement of surge arrestor was scheduled and after “Features and benefits of UHF Partial Discharge Monitoring Systems for replacement, on re-energisation, the PD signal was no longer GIS” Electrical Insulating Materials, 2005. (ISEIM 2005). Proceedings of 2005 International Symposium on. 722 - 725 Vol. 3 August 2005 active.