GRASP Routines: Time-Frequency Methods

0 This section describes the use of software written to extract signals in noise via time frequency methods. Such methods will be useful in extracting signals from poorly modeled or unmodeled sources. The basic procedure we adopt involves the following steps,

• Construction of two dimensional `time-frequency' (TF) maps of the time series data,
• Search for one-dimensional structures in the map.
• Use a statistic based on the length and/or the intensity of the line to determine whether the structure is due to a signal.

Each of the above defined steps can be accomplished by various algorithms. For a detailed discussion of one implementation of this strategy, please refer to [33] and references therein. In this implementation we use the Wigner-Ville distribution to construct the TF map and the Steger's line detection algorithm to detect the line features in the map and a simple length threshold to determine whether we have detected a signal.

In order to improve the computation/communication ratio during code parallelization we introduce the concepts of segment and subsegment. The master process sends out large chunks of data called data segments. Each data segment contains many subsegments and the slave processes compute the TF map for each subsegment in turn. The sizes of the segments and the subsegments are user defined. Also some points at the beginning and the end of each segments are not analysed. This can, of course, be compensated for by padding. The number of points skipped at the beginning and end are termed PRESAFETY and POSTSAFETY respectively and are user defined.