Sound pulse signal processing method for fauld detection

The sound pulse signals recognition for industrial condition monitoring is studied in this paper. The fault feature extraction from signals is the key of condition monitoring, while the general features used for description of pulse signals, such as the slope of rise time curve and power spectral centroid, are incompetent in practice. Aiming at a kind of common pulse signals in industrial production, a new feature parameters - bispectrum weighted value is proposed, in which the frequency components of the signal are given proper weights according to the difference between normal samples and faulty samples. In the experiment of acoustic detection of cracks in the anvil of a large-volume cubic, the feature vector with the proposed parameters results in higher recognition accuracy comparing with the reference feature vector through the diagnosis of BP Neural Network. It has better noise immunity with the help of capability of restraining noise high-order cumulants, and also works well for sound pulse signals disturbed by Gaussian noise in other applications.