Application of various balancing methods to DCNN regarding acoustic data

This paper describes the application and effects of different balancing methods on the learning behaviour and quality of a DCNN using acoustic data. The aim is to show to what extent these methods have positive as well as negative effects on the use case of the audio data. The evaluation is based on synthetic audio data with multiclass characteristics, because an overlay of effects should be avoided. This serves as preliminary work in order to apply the methodology to the measurement data for the classification of knife sharpness in forage harvesters in later investigations. According to applied balancing methods, the data are represented to the DCNN. The performance and quality shall be measured by formal qualification criteria. It turned out that SMOTE gives the best and most robust results. It shows a higher convergence compared to the other methods. Furthermore the worst results are produced with untreated raw data.