Wavelet-based Feature Extraction Methodology for Pattern Classification in Engineering Applications

Feature extraction is fundamental in the framework of pattern recognition. In classification applications where raw data is collected from scientific experiments or simulations, the resulting data may have been exposed to noise, excited by external dynamics, or affected by uncertainties not accounted for during the data acquisition process. Furthermore the data can be high dimensional, which can hinder the accuracy of decision algorithms if additional dimensions of the data make the patterns non-separable, not to mention the added computational complexity burden to train such algorithms. The objective of this research is to understand how specially designed features can improve the class separability in Euclidean space, and how adaptive the feature extraction approach can be when implemented in various pattern classification applications. The wavelet analysis, with known benefits for simultaneous time and frequency localization, has seen recent application to feature extraction in pattern recognition problems. However, there is still a gap in understanding if the entire time-scale domain of wavelets is necessary for pattern classification. Despite the benefits of the additional dimension present in the wavelet domain, there is an added computational complexity for machine learning. Furthermore, information in the wavelet domain that is not useful for separation of classes can in fact degrade the performance of a classifier. This thesis presents a novel filtering approach that can utilize the benefits of wavelets as well as select the most useful regions in the wavelet domain to extract features for improving the classifier performance and thereby reducing the computational complexity. This thesis presents a wavelet-based filtering method for feature extraction. The method extracts optimal regions in the wavelet domain where the between-class separation is improved while the within-class separation is reduced. The approach is verified on four different applications, ranging from electric machine fault diagnosis to medical health detection. The experimental results demonstrate significant improvements in classification accuracies on these applications when using the optimal regions in the wavelets. The approach is also compared with existing methods used on the same applications