FD data

Distribution and power transformers are essential components of any electricity network, hence electrical and mechanical safety of the transformer unit is among the highest concerns of electricity providers. Over the course of their operation, transformers face with a wide range of internal and external disturbances which may lead to a partial or full malfunction of the equipment. The service life and condition requirements for distribution and power transformers are now changed and utilities altered their maintenance policy from time-based to condition-based approach. In fact, the priority has been shifted towards fast and informative assessments with minimum risk to equipment and personnel. Transformers inevitably face with occasional fault and consequent deterioration of the overall performance. Therefore, to maintain the optimal working mode and intermittency of the power supply, transformers undergo a scheduled maintenance works, overhaul, visual inspection, cooling and insulating system tests, and even internal components inspection. Within the past several decades, a wide range of different transformer condition monitoring techniques have been developed.
Among existing diagnostic techniques, the Frequency Response Analysis (FRA) is a well-established testing method known as one of the most accurate and the least invasive evaluation technique. Conventionally, the FRA test results are interpreted via visual comparison of the reference and new measurement, which requires a certain level of experience and credibility from the maintenance personnel. Therefore, for the progress of the FRA interpretation process the application of different approaches have been widely discussed including statistical analysis, machine learning, image processing, to name a few. In this thesis, the proposed smart interpretation of the FRA data entails an automated classification of the transformer working condition along with estimation of the confidence level associated with reported classification results.
A new introduced interpretation technique is applied, formulated, examined, and evaluated in several practical case studies where different scenarios of the winding short circuit and mechanical deformation are emulated on a diverse set of test objects, from a single-phase 1 kVA distribution transformer up to a three-phase 40 MVA power transformer. A new strategy for classification of given test observations is devised and implemented for green, yellow and red operating zones corresponding to healthy, suspicious and critical working conditions, respectively. This strategy is conducted based on the predefined green-to-yellow (healthy-to-suspicious) and yellow-to-red (suspicious-to-critical) decision boundaries represented by critical values of the 12 utilized Statistical Indicators (SIs). Moreover, the classification uncertainty in the form of the confidence level is studied for the first time, and is estimated using bolstered error estimation and bootstrap sampling technique borrowed from the pattern recognition and machine learning. The reported data is of great practical importance since it facilitates a more qualified decision regarding further diagnosis and maintenance of transformer under the test. In summary, this thesis provides a new smart approach to analyze transformer frequency response measurement data, to classify them into normal, suspicious, and critical working conditions, and assign an uncertainty level for each corresponding classification. Having the outcome of this thesis in place, the industrial operators and utility managers would be able to make a more accurate decision on transformer condition