Passive sonar harmonic detection using feature extraction and clustering analysis

A current key problem in the development of passive sonar is the classification of data into its different noise sources. This paper focuses on solving the problem using feature extraction and clustering techniques. The methods described in this paper have been developed for data collected from a single sensor omni-directional passive sonar, with the input data being the extracted frequency tracks from a time-frequency lofagram. A single noise source will exhibit a number of frequency tracks within a lofagram, collectively these tracks form a harmonic set. The problem of harmonic detection is to associate the different components of each of the harmonic sets, or noise sources. Each of the frequency tracks within a harmonic set have a strong physical relationship defining their behaviour and properties. In solving the harmonic detection problem this physical relationship is exploited to associate frequency tracks with similar characteristics. In the solution to this problem each of the frequency tracks present have several key features measured. This allows the characteristics of each of the frequency tracks to be described in a small number of key, directly comparable, parameters. In this paper a small selection of features that may be used in the analysis of the frequency tracks are described. These features are then measured for a number of different data sets. With these sets of parameterised features it is possible to associate harmonically related frequency tracks by implementing clustering analysis, since the strong physical relationship of the features mean that related tracks will be clustered. In applying clustering, decisions need to be made into how many clusters there are in the data. The approach used in this paper uses hierarchical clustering. Hierarchical clustering begins by placing each data point into its own cluster. Two clusters are then merged based on a clustering criterion. At each step two clusters are joined until all the data is held within a single cluster. The progression of the cluster can be shown in a tree diagram, or dendrogram, which is then used to find the optimal level based on the ratio of the distribution of data within a cluster and the separation between different clusters.