Modeling high dimensional frequency spectral data based on virtual sample generation technique

In many situations, such as medical records of rare diseases, early stages of flexible manufacturing system and continuous industrial process, only small training samples can be obtained to construct prediction model. When modelling with high dimensional spectral data, it is very much difficulty to construct efficient and effective prediction model with such a small sample. This research proposes a new virtual sample generation (VSG) approach to model mechanical vibration and acoustic spectra. At first, prior knowledge about the actual training samples is used to produce input of virtual sample. Then, partial least squares (PLS) is used to extract spectral features for reducing features dimension. Thirdly, genetic algorithm (GA) and backup propagation neural networks (BPNN) based feasibility-based programming (FBP) model is used to generate virtual sample´s output. At last, shell vibration and acoustic spectral data of a laboratory-scale ball mill are used to verify performance of the proposed method.