Enhanced damage classification accuracy on a transmission by extending existing datasets with generative adversarial networks

Abstract

In many areas of drive technology, condition monitoring of transmissions and drive systems is becoming an increasingly important discipline. Condition monitoring systems are used in many cases in combination with machine learning algorithms. The generation of a sufficient amount of data per condition class is relevant to ensure training stability and accuracy of the applied algorithms. Especially in early development phases a sufficient data generation is not often given. In the scope of this paper, a Generative Adversarial Network is applied to generate synthetic data and therefore extend existing measurement data sets. Acceleration data in three different condition classes is used, that has been collected on a gearbox as part of the PHM Data Challenge 2009. In order to highlight relevant features and reduce the number of data points, data is pre-processed via appropriate signal analysis techniques, in this case with the spectral kurtosis. It is shown, that in this use case the synthetically generated data via a Generative Adversarial Network has the same feature characteristics as the real measured data sets. The augmentation of the existing data set also improves the detection accuracy with artificial neural networks for the classification of different system states.