Noise/Spike Detection in Vibrations and other Cyclic Signals

The major aim of this study is to describe a unified procedure for detecting noisy segments and spikes in transduced signals with a cyclic but non-stationary periodic nature. According to this procedure, the cycles of the signal (onset and offset locations) are detected. Then, the cycles are clustered into a finite number of groups based on appropriate geometrical- and fre-quency-based time series. Next, the median template of each time series of each cluster is calculated. Afterwards, a correlation-based technique is devised for making a comparison be-tween a test cycle feature and the associated time series of each cluster. Finally, by applying a suitably chosen threshold for the calculated correlation values, a segment is prescribed to be either clean or noisy. As a key merit of this research, the procedure can introduce a decision support for choosing accurately orthogonal-expansion-based filtering or to remove noisy segments. In this paper, the application procedure of the proposed method is comprehensive-ly described by applying it to phonocardiogram (PCG) signals for finding noisy cycles. The database consists of 126 records from several patients of a domestic research station acquired by a 3M Littmann® 3200, 4 KHz sampling frequency electronic stethoscope. By implement-ing the noisy segments detection algorithm with this database, a sensitivity of Se = 91.41% and a positive predictive value, PPV = 92.86% were obtained based on physicians assess-ments.

Keywords:

Filtering , Frequency Time Series , Energy Time Series , Correlation Coefficient , Clustering , Cluster Template , Ensemble Averaging , Orthogonal Reconstruction , Filtering Strategy Curve