The effect of augmentation and filtration on noisy environment's acoustic signals to detect abnormalities in industrial machines based on artificial neural networks

Abstract

The use of acoustic methods for failure detection is increasing in recent times. By using the acoustic method, some features related to the condition of the equipment under test can also be extracted. The main objective of this study was to compare the effect of using filters and augmentation on audio signals in the data training process on the accuracy and loss generated during training and testing. The model is developed using an artificial neural network. The MIMII dataset, which can be accessed openly, is being used for training and testing purposes. Based on the four scenarios designed, it can be seen that the movement of changes in accuracy values in tests involving the use of original data and augmented data is faster (close to 100% after the twentieth epoch) than the other methods. This is because the augmentation process in the form of a time shift does not change the signal in the frequency domain. Moreover, the test signal used is a periodic signal.