Authors:
Andreas Look
;
Oliver Kirschner
and
Stefan Riedelbauch
Affiliation:
Universität Stuttgart, Germany
Keyword(s):
Convolutional Neural Networks, Generative Networks, Condition Monitoring.
Related
Ontology
Subjects/Areas/Topics:
Applications
;
Cardiovascular Imaging and Cardiography
;
Cardiovascular Technologies
;
Health Engineering and Technology Applications
;
Pattern Recognition
;
Sensors and Early Vision
;
Signal Processing
;
Software Engineering
Abstract:
In this paper a convolutional neural network (CNN) with high ability for generalization is build. The task of the
network is to predict the occurrence of cavitation in hydraulic turbines independent from sensor position and
turbine type. The CNN is directly trained on acoustic spectrograms, obtained form acoustic emission sensors
operating in the ultrasonic range. Since gathering training data is expensive, in terms of limiting accessibility
to hydraulic turbines, generative adversarial networks (GAN) are utilized in order to create fake training data.
GANs consist basically of two parts. The first part, the generator, has the task to create fake input data, which
ideally is not distinguishable form real data. The second part, the discriminator, has the task to distinguish
between real and fake data. In this work an Auxiliary Classifier-GAN (AC-GAN) is build. The discriminator
of an AC-GAN has the additional task to predict the class label. After successful training it is possible to
obtain a robust classifier out of the discriminator. The performance of the classifier is evaluated on separate
validation data.
(More)