Robust State Fusers Over Long-Haul Sensor Networks With Applications to Target Tracking

2015-03-01T00:00:00Z (GMT) by Katharine Brigham
<p>In general, sensor networks consist of sensing, data processing, and communication components, and these sensors may communicate with each other or with a central processing center, which then performs some form of data aggregation or data fusion. The terms aggregation and fusion are often used for the same general purpose: how to simultaneously use pieces of information provided by several sources in order to come to a conclusion or a decision. A number of data fusion methods have been developed for sensor networks for a variety of applications, with a primary function of taking in the data from multiple sensors and combining this data to produce a condensed set of meaningful information with the highest possible degree of accuracy and certainty. In this work we primarily explore the use of state fusers for target tracking applications that utilize long-haul communication networks where the underlying target dynamics are nonlinear (as is the case, for example, for a maneuvering target or a ballistic target). However, it is noted that the most popular approaches linearly combine the data. Therefore, the goal of the work is two-fold: 1) investigate/improve nonlinear fusion algorithms for target tracking and 2) develop methods to ensure that these nonlinear fusion algorithms are also robust against packet losses and delays that result from long-haul communications. In particular, we investigate the use of artificial neural networks (ANNs) for multisensor fusion. ANNs possess the capability of modeling arbitrary mappings, as long as a sufficient number of training samples are available from the same distribution. This also provides us with the ability to use nonlinear functions for fusing the data, which may yield better results than with linear fusion given proper training. More specifically, this thesis investigates several aspects of using ANN fusers for multisensor fusion in target tracking. Simulation experiments show that a significant amount of training data is required in close proximity to the test target in order to obtain good performance. Alternate methods in ANN training are then introduced which reduce the amount of training data required to obtain good performance, and widens the allowable training space as well. Then, the use of multiple fusers, different input features, and varied ANN architectures are investigated with the intent to further improve fuser performance. The effects of imperfect communications are then explored for the ANN fuser, and another training enhancement is suggested to generate ANN fusers that are more robust against packet losses and delays. Overall, this thesis intends to provide suggestions as to what parameters or aspects of the ANN may be explored to help improve fuser performance for use in target tracking.</p>