Anomaly Prediction in Non-Stationary Signals using Neural Network Based Multi-Perspective Analysis

A new technique for predicting anomalies in the near future of an observed signal is being presented. Before any data analysis can be performed on an observed signal, the signal's underlying pattern must be cleared. A wavelet de-noising scheme is used because it provides a better result compared to other de-noising algorithms and it is simple from a computational standpoint. Robust peak-finding algorithm is used to find smaller anomalies that appear frequently throughout the signal pattern. In addition to or in place of wavelet de-noising, other views of the signal may be generated for analysis. The generated perspectives are used as input to a feed-forward neural network that will predict the likelihood of an anomalous event occurring later in the signal. The neural network is trained using the Resilient Backpropagation of Errors (Rprop) supervised learning algorithm with data sets consisting of a mix of signals known to precede anomalous events as well as signals known to be free of significant anomalies. This paper provides a means of predicting large or abnormal events in signals such as seismograms, EKGs, EEGs, and other non-stationary signals. Our algorithm has been tested on a large collection of seismic and EKG (electrocardiogram) signals. The obtained accuracy as high as 70% with EKG signals and as high as 83% with seismic signals, when the test data is taken from within the same time frame as the training set. Though there was greater consistency found at a lower degree of accuracy for seismic signals. [Abdullah Alshehri, Aaron Waibel, Soundararajan Ezekiel. Anomaly Prediction in Non-Stationary Signals using Neural Network Based Multi-Perspective Analysis. Life Sci J 2014; 11(6):685-693]. (ISSN: 1097-8135). http://www.lifesciencesite.com. 104