Please use this identifier to cite or link to this item:
|Title:||Neural Networks for Data Fusion|
|Department:||Electrical and Computer Engineering|
|Keywords:||Electrical and Computer Engineering;Electrical and Computer Engineering|
|Abstract:||<p>The process of data fusion may be viewed as a multi-level hierarchical inference process whose ultimate goal is to assess a mission situation and identify, localize and analyze threats. Each succeeding level of data fusion processing deals with a greater level of abstraction. The lower level is concerned solely with individual objects where sensor data are processed to derive the best estimates of current and future positions for each hypothesized object as well as provide an inference as to the identity and key attributes of the objects. With the recent proliferation and increasing sophistication of new technologies. It is recognized that the incorporation of new techniques, such as neural networks and others, will make the data fusion system more powerful in tri-service (command, control, and communications). In this thesis, optimization neural networks are investigated. A new technique of measurement data association in a multi-target radar environment is developed. The technique is based on the mean-field-theory machine and has the advantages of both the Hopfield network and the Boltzmann machine. In the technical development, three new energy functions have been created. Theoretically, the critical annealing temperature is found to determine the annealing temperature range. A convergence theorem for the mean-field-theory machine is put forward. Based on the technique, neural data association capacities have been evaluated in cases with and without clutter, taking into account different accuracies for radar measurements. New energy functions have been extended to multiple dimensional data association. A comprehensive analysis by computer simulations has demonstrated that the new technique developed here possesses high association capacity in the presence of false alarms; it can cope with track-crossing in a dense target environment. A feature-mapping neural network for centralized data fusion is presented, and its performance is compared with that of the Maximum Likelihood approach. In support of our study of multisensor data fusion for airborne target classification with artificial neural networks (ANNs), we designed a neural classifier. Multilayer perceptrons neural networks trained by back-propagation (BP) rule are discussed. In order to speed up the training or decrease the number of epoches in the learning process, both momentum and adaptive learning rate methods are used. The simulation results show that the technique of automatic target classification using neural networks has the potential to classify targets.</p>|
|Appears in Collections:||Open Access Dissertations and Theses|
Items in MacSphere are protected by copyright, with all rights reserved, unless otherwise indicated.