Automatic Partial Discharge Classification of Gas-Insulated Systems at DC Voltage

This thesis deals with the automatic identification and classification of typical partial discharge defects of gas-insulated DC systems using various innovative machine learning algorithms. The aim is to develop a robust partial discharge classification system for gas-insulated switchgears (GIS) at DC voltage.

Starting with a detailed introduction to the topic of machine and deep learning. Practical examples are presented in which artificial intelligence is already used in high-voltage engineering applications. The algorithms, methods, and strategies from the field of machine learning used in this work are described in detail and the mathematical principles are discussed to provide a deeper understanding. Most algorithms require characteristic parameters that are used to achieve a correct classification. These values are referred to as features and serve as input parameters for the models. The selection of the correct features is, therefore, of great importance. In addition, the main focus is thus on the extraction, engineering, and analysis of the features as well as the final model selection including validation strategies and interpretation of the results. Methods and strategies from the literature are presented for this purpose. Moreover, an overview of statistics is given, as statistical parameters to describe the partial discharge defects at DC voltage are the basis of this work.

In order to realise a reliable automatic partial discharge classification system for gas-insulated DC systems, the requirements and structure of a modern partial discharge monitoring system for GIS are discussed. The most important software features of this system can be divided into the separation, identification, and risk assessment of the detected partial discharges. The scope of this work focuses on the first two points. Finally, a proposal for the structure of such an automatic partial discharge classification system used in a monitoring system for DC GIS is developed.

On this basis, the investigated partial discharge defects are analysed in detail in order to find characteristic features that allow a clear distinction between each defect type. The distribution functions of the measured variables and parameters derived from them were used for this purpose. In a next step, the features are then extracted from the data. The database used for this comprises about 1,900 different measurement files and more than 18,000 PD pulse sequences. To gain a deeper understanding of the extracted characteristics, an exploratory data analysis was carried out.

Based on the previously gained knowledge and results, various machine learning algorithms were trained, validated and tested. For this purpose, electrical partial discharge measurements under SF6 were used, whereby high classification accuracies were achieved. In addition, the models obtained were tested with measurement data from ultra-high frequency sensors in order to evaluate the robustness of the selected features and the applicability of these in DC GIS.