Motivation
Cavitation occurs in liquids when pressure locally drops below vapour pressure. When such created vapour bubbles enter regions of higher pressure, they will implode. This implosion can cause damage, if the implosion happens near solid surfaces. The process of nucleation, growth, and implosion is called cavitation.
Since visual inspection of cavitation events in hydro power plants is not possible, due to missing optical accessibility, acoustic event detection offers an alternative. The radiated cavitation noise is recorded by acoustic emission sensors operating in the frequency range between 100kHz and 1MHz. Below 100kHz the sensors behave like high-pass filters. The upper frequency bound is limited by the maximum sampling frequency of 2MHz.
Data Preprocessing
In the first step the signal is divided into 0.1s snippets. Afterwards short-time Fourier transformation (STFT) is applied. Since a wide frequency range shall be analyzed, dynamic range compression (DCR) is used in order to reduce dynamics between high and low frequencies. For dimensional reduction the frequency axis is wrapped by a filter bank, which consists of 32 equally spaced and non overlapping triangular windows. In the last step each spectrogram is scaled between 0 and 1.
Conventional Acoustic Event Detection
For conventional acoustic event detection a convolutional neural network with three convolutional layers and three dense layers is used. Neural Networks represent a class of Algorithms for pattern recognition applications, which try to mimic the human brain.
Dropout and max-pooling is used in order to increase generalization capability. Using early stopping a categorical accuracy of 94.2% can be achieved. The accuracy was evaluated at separate power stations.
Generative Adversarial Networks
Generative adversarial neural networks (GANs) are capable of synthesizing realistic pictures/ sound files. In every GAN-structure two networks compete against each other. The first network (generator) synthesizes files and the second network (discriminator) has to distinguish between real and fake samples. It is possible to synthesize spectrograms and to use these as additional input in order to train a classifier. Using an AC-GAN (Figure 3) the discriminator can later be directly used as a classifier.
Using an AC-GAN the categorical accuracy could be raised to 98%. Furthermore using this semi supervised training method comes with the advantage of an declined variance of categorical accuracy.
