The U.S. Navy seeks to develop techniques that improve the exploitation of ephemeral features in active and passive sonars. Off-the-shelf classifiers that use multiple hypothesis and machine learning tools are limited in their ability to handle ephemeral features, as they expect the feature availability all the time. SSC proposes to use novel sparsity-aware split finding methodology on sonar data. The innovative techniques not only accommodate ephemeral data but maintain the purity metrics associated with the data allowing for highly accurate classification. In Phase I, these techniques are demonstrated on data collected from in-air tests but can easily be generalized to Navy operations. These techniques will be extended to sonar data sets in Phase II. This will improve identification and tracking of stealthy targets and reduce time to react. The overall methodology is widely applicable to operational settings for various industrial uses and could see wide commercial applicability.
Benefit: The novel sparsity-aware split finding techniques could be generalized and find widespread commercial use, such as in fault detection and identification, when the machinery is operated in a variety of modes and with different operational objectives. They could also be applied to non-destructive testing, condition-based maintenance, and remote monitoring of equipment, which can have great commercial value.
Keywords: Decision handling, Decision handling, Software development, Novel training methods, Sonar classification algorithms, Ephemeral features, Surface ship sonar, Signal processing, Neural Network classifiers
