Model-based data augmentation to improve the performance of machine-learning diagnostic systems

IET conference proceedings., vol. 2021, pp. 545–550

Abstract

Machine-learning diagnostic systems are already being used to detect abnormal conditions in electrical equipment. The challenge in machine-learning diagnostic approaches lies in dealing with small training databases of abnormal-condition samples available as the performance of the machine-learning diagnostic systems relies on the quality and quantity of the training data. One possible solution to this problem are data augmentation techniques generating synthetic data from the available data to augment the training data. However, these generic methods do not consider domain-specific knowledge about the diagnosis task. In this paper, a model-based data augmentation approach using computer-implementable, electromechanical models is presented. This approach uses statistical information extracted from the available data to sample model parameters and input variables to generate synthetic normal- and abnormal-condition data. The model-based data augmentation is showcased for detecting a distribution transformer fault. Machine-learning diagnostic systems are trained utilizing the model-based data augmentation and benchmarked with state-of-the-art diagnostic systems. The number of abnormal-condition measurements for training is limited and the performance of all diagnostic systems is analysed. It is shown that the model-based data augmentation is capable of improving the diagnosis accuracy for the analysed diagnostic task especially when only a very small amount of abnormal-condition data is available.

Authors 3

1. Jannis Nikolas Kahlen Aachen

   RWTH Aachen University · Fraunhofer Institute for Applied Information Technology

   Affiliation as printed

   Fraunhofer Institute for Applied Information Technology FIT, Digital Energy, Schloss Birlinghoven 1, Sankt Augustin, Germany

   RWTH Aachen University, Schinkelstraβe 6, 52062 Aachen, Germany

   Fraunhofer Institute for Applied Information Technology FIT, Digital Energy, Schloss Birlinghoven 1 , Sankt Augustin , Germany

2. Andre Würde Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Schinkelstraβe 6, 52062 Aachen, Germany

3. Albert Moser Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Schinkelstraβe 6, 52062 Aachen, Germany