| In this work we achieved several goals. First a new technique for organizing signal recognition using a Tier structure is proposed. The underlying philosophy of this structure has two purposes; one, to organize the signal features based on their complexity and two, to provide a means by which evolutionary techniques can evaluate optimal combinations of features for the given problem. The approach provides layers for organizing signal recognition, Tier I consisting of time-frequency features, Tier II, contour (or scalogram) features, and Tier III single point application features. The Tier structure is combined with a four-stage recognition process. The four stage process consists of feature extraction (stage 1), unsupervised learning (stage 2), detection (stage 3) and performance evaluation (stage 4). The signal recognition technique is applied to analyzing web vibrations made by a jumping spider, Portia fimbriata. For this application Tier I and II level, Continuous Wavelet Transforms (CWT) are used, whereby contour data is extracted. Application specific features are developed at the Tier III layer. All features are managed through a computer-aided tool, CLARA, developed specifically for this research. The multi-resolution Tier approach successfully detected, for the first time, subtle but significant differences in the Portia "smokescreen" signal. Results indicate significant improvement on detection for this application over previous approaches. Lastly a fifth stage is added to the signal recognition system, which couples optimization with the Tier architecture. A process is proposed which combines the recognition architecture with an evolutionary based algorithm intended for feature mining. For this work, the author points out the challenges associated with signal recognition, referring to the vast number of feature parameters that can be derived at the Tier I, II and III layers. Considerable time is spent on developing parameter settings for Tier I layer, where Short-Time Fourier Transforms (STFT) and CWT based approaches can be interchangeably used. Chromosome coding along with a fitness relation based on the receiver operator curve is discussed. |
