Pattern Recognition in Signals Through Rough Concept Bounding

The recognition of complex signal patterns is generally considered to be a difficult task. Analytical approaches and rule-based expert systems have been used for the recognition of signals or waveforms. Using an analytical approach requires mathematical training and substantial knowledge of the particular details of the waveforms; hence the development process of a recognition system is often time-consuming. Rule-based approaches also require a knowledge engineer to work closely with a domain expert in order to formulate the rules for recognition. Because of these limitations, it is highly desirable to formulate a methodology to extract the representative features of signal patterns and use them directly in the recognition process. Adaptive signal decomposition is one procedure that can be used to extract relevant features from the signal segments. The results of the decomposition are time-frequency atoms from a common vocabulary. Due to noise and the context embedding the signal segment, however, variation occurs in the decomposed results produced for very similar signal segments as well as in the number of time-frequency atoms acquired. Thus, the target signal segment can only be compared roughly with the learned example signal. In the present study, rough comparison methodology was developed, on the basis of the rough concept, to measure the similarity between the signal segments. A signal recognition system was constructed incorporating adaptive signal decomposition and rough comparison strategies, allowing the system to learn exemplars from different classes and to recognize and classify the target signals during the search through a stream of signals. The rough comparison approach employs the notion of lower and upper approximation signatures for each signal class. The lower approximation signature (LAS) consists of the necessary features for the class; the upper approximation signature (UAS) encompasses all the possible features that the exemplars in the class may have. During learning phase, the feature weights of the LAS and UAS of each class are adjusted. Each feature weight is specified as a function of feature, class, and approximation type. On the basis of the learned features, target signal segments can be recognized. In both the learning and recognition processes, the rough comparison approach is used to measure the similarity between the signal segments.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2015

Author dc:creator

• Gyaw, Tun Aung

Contributors dc:contributor

• Sylvian R. Ray

• Engineering, Biomedical

Language dc:language

eng

Identifier

(MiAaPQ)AAI9834683

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/81908