Fully Interpretable Optimized Weights-Based Methodology for Adaptive Machine Fault Components Extraction

Rotating machines are widely used in various domains and its condition monitoring can help to gain more economic profits and prevent unexpected accidents. Vibration signals usually contain sufficient health information, but they are also severely contaminated by heavy background noises including random noise and fixed fundamental vibration components. Therefore, it is of vital importance to effectively extract fault components for effective machine condition monitoring. Considering that the fault components of rotating machines usually exist in some narrow frequency bands, blind bandpass filtering methods such as fast kurtogram, blind deconvolution, and adaptive signal decomposition have been successively proposed. However, it is observed that these methods might be easily influenced by random impulsive noise and low-frequency components. Moreover, these methods also cannot simultaneously extract fault components distributed in different frequency bands. To solve these problems, a new machine fault components extraction methodology based on fully interpretable optimized weights is proposed. The interpretable optimized weights are generated by convex optimization by using healthy and faulty vibration signals, and they could provide sufficient frequency information of fault components. Simulated and experimental vibration signals demonstrate the effectiveness and superiority of the proposed method in solving the aforementioned problems.