Cavitation detection in hydraulic machinery by analyzing acoustic emissions under strong domain shifts using neural networks

We propose a novel, general-purpose framework for cavitation detection in a wide variety of hydraulic machineries by analyzing their acoustic emissions with convolutional neural networks. The superiority of our system lies in the fact that it is trained exclusively with data from model turbines operated in laboratories and can directly be applied to different prototype turbines in hydro-power plants. The challenge is that the measurements to train and test the neural network stem from machines with various turbine designs. This results in train and test data with different data distributions, so-called multi-source and multi-target domains. To handle these domain shifts, two core methods are provided. First, an advanced pre-processing pipeline is used to narrow the domain shift between data from different machines. Second, a domain-alignment method for training neural networks under domain shifts is used, resulting in a classifier that generalizes well to a wide range of prototypes. The outcome of this work is a generic framework capable of detecting cavitation in a wide range of applications. We explicitly do not try to obtain the highest accuracy on a single machine, but rather to achieve as high as possible accuracy on many machines.