Heterogeneous Data Discovery Using Deep Neural Networks

Improving feature extraction, event detection, and target classification in multi-sensor systems requires new mathematical methods and processing techniques. In addition, previous research and experience suggests that leveraging sensor data that has not experienced significant dimensionality reduction can preserve subtle features when processed jointly with other relevant data. However, traditional ISR solutions typically process and analyze sensor outputs independently, often without a priori, or jointly-coupled, information between other sensor types - leaving the capabilities of these systems limited and predictable. Therefore, we propose research to discover new multi-sensor methods, and overcome traditional single-sensor limitations, by utilizing and developing state-of-the-art nonlinear processing methods based on the fusion of Deep Convolutional Neural Networks (DCNN) and non-parametric signal processing algorithms. We will combine new mathematical methods and robust processing techniques in order to discover new observables from multiple inputs, even from sensors that return disparate forms of data. The proposed methods are will suited to exploit jointly-coupled information from multi-sensor data that is analyzed up-stream (i.e., in a raw or weakly-processed form). This research will open up a new area for continued research along with the opportunity to discover new features in multi-sensor configurations that are not available in traditional stove-pipe solutions.