Intelligent partial discharge evaluation of generator insulators using machine learning methods (SmartPD)
Recognizing partial discharge patterns in electrical machines:
The service life of electrical machines substantially depends on the condition of the insulation. This insulation is assessed manually based on partial discharge measurements. These measurements show the current PD behavior. The trained eye can recognize PD patterns and infer the causes of PD. Statements on the service life are only possible by performing regularly repeated measurements. In large stationary generators, automated measurements are particularly beneficial for the early detection of failures.
In the SmartPD project, algorithms are tested and trained to recognize these PD patterns. Error patterns are measured specifically for training the algorithm. This data is then used for training and testing an algorithm based on machine learning.
The project is interesting to:
Manufacturers of electrical machines, providers of monitoring technology / partial discharge measurement technology
Developing a classification method for partial discharge patters in electrical machines
Evaluating the method for both offline and online PD measurements
Neural networks, MLP, CNN, data augmentation
Detailed project description
As part of the project, exemplary PD measurements on electrical winding designs were initially performed in collaboration with the University of Kassel. In addition to the measurements, the most common methods for measuring partial discharge and their forms of presentation were investigated. Based on these measurements and the results of the investigation, the decision was made to use exclusively PRPD patterns as input data for the classification. This form of presentation is supported by all common partial discharge measurement systems.
As it is not possible to depict all error patterns based on the electrical winding designs, the error patterns were obtained from the IEC TS 60034-27-1:2012 and IEC TS 60034-27-2:2012 standards in the course of the project. To generate a sufficient number of training patterns for a CNN classifier, 400 patterns were generated for each error type on the basis of the normal patterns. To create sufficient diversity in the patterns, they had to be generated by stretching and compression and by adjusting the black / white distribution for each row of pixels of a pattern.
The patterns generated that way are used to train a CNN network that assigns the patterns to a specific error type. Two different label categories were examined as target parameters. The first category only roughly classifies the error into three error types: slot discharge, internal partial discharge, and partial discharge in the end windings. The second category allows a more precise error classification into a total of seven error types.
Generating PD patterns at the test station
Selecting, generating and preparing training data
Developing and evaluating the models for the classification of PD patterns
Participants from Fraunhofer IEE:
Sebastian Lengsfeld, Florian Rehwald, Hardy Ast, Constantin Arronsohn, Bela Schinke
A publication as part of the ICEM 2022 Valencia has been submitted.