Files and links (1)
PDFOpen Access (License Unspecified), Open Access
Dissertation
Pretraining and shelving of neural networks by Fourier analysis and their use for target classification
Doctor of Philosophy (Ph.D.), Drexel University
Dec 1993
DOI:
https://doi.org/10.17918/00008687
Abstract
This effort was motivated by a problem of present and practical interest, namely, the classification of targets using radar data. The multi layer perceptron (MLP) neural network and its associated Back Propagation Algorithm (BPA) were proposed as a solution because of the expected payoff in robustness to some radar parameters such as dwell time. In so doing, the limitations associated with the BPA, such as the well documented long training time, had to be dealt with. The method proposed to reduce the training time of the BPA partitioned the training process in 2 steps, pretraining and final training. A network was first pretrained using one of four approaches. Final training then consisted of: (1) changing the topology of the pretrained network to accommodate the requirements of the specific target classification problem and; (2) training the network to classify the targets using the BPA. The final training step departed from the typical practice of initializing the weights of the MLP randomly. Instead, initial weights for the final training were obtained at the end of the pretraining process. The pretraining process trained a network to recognize a set of basic signal components that form the target's signature. If the basic signal components are selected independent of any specific target set, then the pretrained network can be shelved for future use against different target classification problems. For the specific problem of target classification by radar, the Fourier basis was proposed as the set of basic signal components. components. The objective of the pretraining process was to train the network to recognize sinusoids. Four pretraining approaches were developed to perform the spectrum estimation function. The algorithms for two of those approaches were tested against simulated signals. One of the two algorithms was used to classify aircraft using radar data. The performance of that one approach was compared to the performances of the baseline BPA and of an enhanced baseline that augmented the BPA by initializing the weights in a different fashion. The results indicate that the pretraining approach was effective at reducing the training time. Other benefits were also noted.
Metrics
35 File views/ downloads
17 Record Views
Details
- Title
- Pretraining and shelving of neural networks by Fourier analysis and their use for target classification
- Creators
- Gerardo J. Melendez
- Contributors
- Stanislav B. Kesler (Advisor) - Drexel University, Drexel University (1970-)
- Awarding Institution
- Drexel University
- Degree Awarded
- Doctor of Philosophy (Ph.D.)
- Publisher
- Drexel University; Philadelphia, Pennsylvania
- Number of pages
- xv, 188 pages
- Resource Type
- Dissertation
- Language
- English
- Academic Unit
- College of Engineering (1970-2026); Drexel University
- Other Identifier
- 991021889115504721