Machine Learning is the study of methods for programming computers to learn. Computers are applied to a wide range of tasks, and for most of these it is relatively easy for programmers to design and implement the necessary software. However, there are many tasks for which this is difficult or impossible. These can be divided into four general categories. First, there are problems for which there exist no human experts. For example, in modern automated manufacturing facilities, there is a need to predict machine failures before they occur by analyzing sensor readings. Because the machines are new, there are no human experts who can be interviewed by a programmer to provide the knowledge necessary to build a computer system. A machine learning system can study recorded data and subsequent machine failures and learn prediction rules. Second, there are problems where human experts exist, but where they are unable to explain their expertise. This is the case in many perceptual tasks, such as speech recognition, hand-writing recognition, and natural language understanding. Virtually all humans exhibit expert-level abilities on these tasks, but none of them can describe the detailed steps that they follow as they perform them. Fortunately, humans can provide machines with examples of the inputs and correct outputs for these tasks, so machine learning algorithms can learn to map the inputs to the outputs. Third, there are problems where phenomena are changing rapidly. In finance, for example, people would like to predict the future behavior of the stock market, of consumer purchases, or of exchange rates. These behaviors change frequently, so that even if a programmer could construct a good predictive computer program, it would need to be rewritten frequently. A learning program can relieve the programmer of this burden by constantly modifying and tuning a set of learned prediction rules. Fourth, there are applications that need to be customized for each computer user separately. Consider, for example, a program to filter unwanted electronic mail messages. Different users will need different filters. It is unreasonable to expect each user to program his or her own rules, and it is infeasible to provide every user with a software engineer to keep the rules up-to-date. A machine learning system can learn which mail messages the user rejects and maintain the filtering rules automatically. Machine learning addresses many of the same research questions as the fields of statistics, data mining, and psychology, but with differences of emphasis. Statistics focuses on understanding the phenomena that have generated the data, often with the goal of testing different hypotheses about those phenomena. Data mining seeks to find patterns in the data that are understandable by people. Psychological studies of human learning aspire to understand the mechanisms underlying the various learning behaviors exhibited by people (concept learning, skill acquisition, strategy change, etc.).

Machine learning

As transformers age, their internal condition degrades, which increases the risk of failure. To prevent these failures and to maintain transformers in good operating condition is a very important issue for utilities. Traditionally, routine preventative maintenance programs combined with regular testing were used. The change to condition-based maintenance has resulted in the reduction, or even elimination, of routine time-based maintenance. Instead of doing maintenance at a regular interval, maintenance is only carried out if the condition of the equipment requires it. Hence, there is an increasing need for better nonintrusive diagnostic and monitoring tools to assess the internal condition of the transformers. If there is a problem, the transformer can then be repaired or replaced before it fails. An extensive review is given of diagnostic and monitoring tests, and equipment available that assess the condition of power transformers and provide an early warning of potential failure.

Review of condition assessment of power transformers in service

Pattern recognition

Dissolved gas analysis (DGA) is used to assess the condition of power transformers. It uses the concentrations of various gases dissolved in the transformer oil due to decomposition of the oil and paper insulation. DGA has gained worldwide acceptance as a method for the detection of incipient faults in transformers.

A review of dissolved gas analysis measurement and interpretation techniques

A concise review is given of the progress made in the field of partial discharges (PD) at DC voltage. Although ample reference will be made to work of other authors in this field, the paper will concentrate on the progress that was made at Delft University of Technology over a period of 14 years in three PhD projects. In the first project, a start was made with the analysis of the physics of partial discharges at DC voltage and different types of PD were characterized based on parameters like time interval between PD and PD magnitude. In a second project, PD analysis was applied to HVDC apparatus and different means of classification of PD at DC voltage were proposed. In the third project, PD analysis was applied to HVDC mass-impregnated cables and test specifications were proposed. In this paper the work performed in the above three Ph.D. projects is summarized with ample reference to papers of other workers in this field. Attention is given to the mechanism of PD at DC voltage as compared to AC voltage and techniques for measurement and analysis of DC PD patterns. Examples of practical application of DC PD testing are given. Finally, some thoughts on future work are presented.

/pdf/partial-discharges-at-dc-voltage-their-mechanism-detection-2ka0gekh8r.pdf

Partial discharges at DC voltage: their mechanism, detection and analysis

Partial discharge (PD) measurement and identification have great importance to condition monitoring of power transformers. In this paper, a new method for recognition of single and multi-source of PD based on extraction of high level image features has been introduced. A database, involving 365 samples of phase-resolved PD (PRPD) data, is developed by measurement carried out on transformer artificial defect models (having different sizes of defect) under a specific applied voltage, to be used for proposed algorithm validation. In the first step, each set of PRPD data is converted into grayscale images to represent different PD defects. Two “image feature extraction” methods, the Local Binary Pattern (LBP), and the Histogram of Oriented Gradient (HOG), are employed to extract features from the obtained gray scale images. Different variants of Support Vector Machine (SVM) are adjusted for optimal classification of PD sources in this process. Impact of the employed parameters in the image processing such as image resolution, random noise, and phase shift, on identification accuracy is investigated and addressed. It is shown that by using HOG-SVM method 99.3% accuracy can be achieved. This is hardly affected by various external factors. Two case studies have been conducted on multi-source PD for evaluating the performance of the proposed algorithm. A void defect is implemented into the transformer model and the resultant recorded signal is used for the study. The DBSCAN algorithm is used as the mean of PD source clustering and sub-PRPD pattern development. It is shown that HOG-SVM method has superior performance in identifying active sources, under sub-PRPD pattern application.

Partial Discharges Pattern Recognition of Transformer Defect Model by LBP & HOG Features

For pt.I see ibid., p.3-14, (2007). Insulation assessment of HV cables requires continuous partial discharge (PD) monitoring to identify the nature of insulation defects and to determine any degradation trends. However to recover PD signals with sufficient sensitivity to determine such insulation degradation in substations with high levels of electromagnetic interference is a major challenge. This paper is the second of two papers addressing this challenge for on-line PD measurements in a noisy environment. The first paper described a wavelet transform-based method of interference rejection. This paper applies that method to the problem of on-site testing, using both laboratory tests and on-site tests. The laboratory tests were used to stimulate the noisy on-site testing environment, with use of transient pulse-like noise, discrete spectral interference (DSI) and white noise. These noise types have been successfully rejected by the method proposed in the first paper. In addition, on-site tests have been undertaken and have been able to detect PD signals in an old 11 kV substation multi-cable installation

A novel wavelet transform technique for on-line partial discharge measurements. 2. On-site noise rejection application

Frequency Response Analysis (FRA) has been in use since the last decade as a sensitive method for detecting transformer internal defects. This study is focused on the influence of temperature and moisture migration on the FRA response of transformer winding. It also discusses the feasibility of FRA capability in moisture diffusion recognition in transformer paper insulation. To conduct this study, a single phase model transformer involving concentric LV and HV windings and a 20/0.4 kV, 1.6 MVA threephase two windings transformer are taken as test objects. Experiments are carried out at different temperature and moisture conditions. FRA spectra are then recorded and analyzed. FRA spectra deviations as well as total capacitance variations due to the temperature and moisture changes in the test objects are calculated. Karl-Fischer Titration (KFT) is utilized to monitor the moisture migration within oil and paper insulations. Furthermore, a mathematical model is used to simulate one of the test object windings and verify the experimental result, and also clarify the main reason of FRA spectrum deviation in this circumstance. Finally, statistical indices in FRA evaluation are calculated to explore their capability in FRA spectrum interpretation once the moisture content of paper insulation is changed.

Influence of temperature and moisture content on frequency response analysis of transformer winding

High impedance fault (HIF) has long been a challenging problem in network protection due to its random behaviour and low magnitude. The conventional protection devices cannot be utilised since the HIF does not draw enough current to cause tripping. The feature of wavelet transform (WT) which decomposes a signal into different frequency bands and locations in time can be utilised to extract HIF features and detect its occurrence. In the study, HIF arcing currents associated with different types of contact surface are produced by experiments set up in the laboratory. Features of the fault currents are extracted and a new detection criterion is developed based on WT coefficients. Transformer magnetising inrush currents and capacitor switching transients are also produced by experiment and simulation, respectively, to represent transient disturbances commonly occurred in the distribution network. The detection criterion is used to discriminate the HIF from other normal non-fault transient events. Three different types of wavelet are tested and the Daubechies wavelet Db4 gives the best performance based on its detection and discrimination rates. The efficacy of utilising existing conventional current transformers for HIF application is also tested and verified by experiment.

https://ietresearch.onlinelibrary.wiley.com/doi/pdf/10.1049/iet-gtd.2016.0021

Detection of high impedance faults using current transformers for sensing and identification based on features extracted using wavelet transform

Innovations in computer technology have made possible continuous on-line monitoring of partial discharge (PD) activities. The power industry aims to assess the condition of power system equipment through on-line monitoring of PD activities. This involves long-term continuous data recording and it is very difficult to extract useful information from such a large amount of raw data, particularly if it is done manually. Instead, data mining can be applied in solving this problem. Data mining can be categorized into predictive modelling and descriptive modelling. In this paper, work was mainly focused on predictive data mining, which is classification of PD. The back propagation neural network (BPN), self-organizing map (SOM) and support vector machine (SVM) were used for classification and compared. Results indicate SVM is the best method in terms of classification accuracy and processing speed.

Trevor Blackburn

Papers

Partial Discharges Pattern Recognition of Transformer Defect Model by LBP & HOG Features

A novel wavelet transform technique for on-line partial discharge measurements. 2. On-site noise rejection application

Influence of temperature and moisture content on frequency response analysis of transformer winding

Detection of high impedance faults using current transformers for sensing and identification based on features extracted using wavelet transform

Application of data mining on partial discharge part I: predictive modelling classification