Optimal Frequency Band Selection Based on the Clustering of Spatial Probability Density Function of Time-Frequency Decomposed Signal

Heavy-duty machines are often working in the harsh conditions. Components of such machine will suffer significantly higher stress and will tend to wear off more quickly. Thus, it is essential to detect fault in its early stages. One of the methods of detection is selection of the optimal frequency band (OFB) for the filter design. One can find such filter characteristic through statistical approach or iterative or adaptive methods. Authors in their research propose to use time-frequency decomposition via STFT as it is one of the quickest algorithms to apply to the data. However, due to the signals structure, there will be high energy in the lower frequency band. Thus, it is reasonable to perform normalization of the absolute value of STFT matrix. Knowledge of the signals structure allows us to distinct three main different signal components. First component is a noise, usually Gaussian, second - impulsive behavior related to the fault and last one - accidental high energy impacts which disturb performance of most of the algorithms. Therefore, authors propose to model each of the sub-signals of the decomposed signal with probability density functions (PDF). Different components will give different PDF. In the final step, authors have proposed to use k-means clustering algorithms to distinguish between different structure of the frequency bands and select optimal one for the filter characteristic design.