Feature extraction and selection based on vibration spectrum with application to mill load modeling

Feature extraction and selection were important issues in soft sensing. Using the spectrum of vibration or acoustical signal may simplify the modeling process. In this study, shell vibration signals of ball mill were first transformed into vibration spectrum by fast Fourier transform (FFT). Then, the candidate features set were extracted from the spectrum, which included three types of features: characteristic frequency sub-bands, spectral kernel principal components (KPCs), masses and central frequencies of spectral peaks. We used several techniques, such as genetic algorithm (GA), partial least square (PLS) and kernel principal component analysis (KPCA), to obtain these features. The optimal selection of input sub-features and model parameters were calculated by GA based optimization method. The test results showed that the proposed approaches were effective for modeling parameters of mill load.