A feature-based transient signal detection methodology

A neurofuzzy transient detection methodology is developed and shown to solve transient detection problems which abound in many applications. Unknown additive noise type, low signal to noise ratio as well as shortage of signal information often complicate the problem of transient detection. Many conventional methods have been developed and applied to this problem. However, they are found to be effective in limited situations and typically they require explicit mathematical models, making them less portable and sometimes not robust. To enhance the new methodology and test its efficacy, several features are investigated including, but not limited to, statistical features, energy of frequency components, and wavelet coefficients. The information they provide is fused via a fuzzy system. The truth decision curves, which are the critical part in the fuzzy system, are obtained by artificial neural networks. In this thesis, the effectiveness of the new methodology is verified through several experiments and comparisons with conventional approaches using both simulated and actual data. The results indicate that the new methodology is capable of detecting transients accurately; of identifying trends reliably; and of not misinterpreting a steady state signal for a transient one. Several applications of the feature-based methodology are also presented. Transient detection improves the accuracy of forecasting systems. To test the efficacy of the new transient detection methodology, it has been integrated in an electric load forecasting system, in which artificial neural networks are utilized as forecasting tools and the transient detector is used to detect the onset of load change. Load predictions for a variety of power customers and isolated power units are obtained. The results show that the new methodology can help power load predictors overcome many difficult problems and can significantly enhance the accuracy of the prediction