Feature extraction of power disturbance signal using time frequency analysis

Abstract

Power Quality has been one of the great concerns recently; it due to the increasing number of loads which sensitive to the power disturbance. One of the main issues in power quality problems includes how to localize each disturbance event and recognize its respective type of disturbance more efficiently. Another problem is harmonics problem which is due to nonlinear loads and the source of fault is difficult to detect and diagnose. Thus, it is important to propose an effective feature extraction method in order to build a system with DSP approach to overcome this problem as well as to maintain the power quality. This thesis utilized the concepts of time frequency analysis (TFA), which provides information of the disturbance signal as a function of time and frequency in order to analyze the power disturbance signals due to those signals is finite energy or non-stationary signals. By choosing real and simulated power signals, this study has been carried out over 30 normal signals and 90 signals with power disturbance including sag, swell, interruption, harmonics, transient and frequency variation. Those signals are transformed into time frequency plane using Bdistribution algorithm. Then the important feature vectors or components are extracted using Singular Value Decomposition (SVD) and Principle Component Analysis (PCA). Finally, the distance metric, J, as class separibity between two classes of vectors can be measured using Maximum Margin Criterion (MMC). From the results obtained, the most two of the right singular vector (SVs) become most powerful feature vectors to describe the TFD. The lowest SVs have cyclic structure becomes less significant feature vector which contains noise or redundancy. Furthermore, the projection between two SVs of normal power signal and disturbance power signal shows the plotting of these vectors are overlap or not overlap respectively. If the last two SVs (or either one is last SV) are projected, the plotting almost approached to zero. The most discriminates vectors is the distance between them, MMC shown either that vectors are close to those in the same class (ranges of J is 0.0006 to 0.0045) or far from those in different classes (ranges of J is 0.0045 to 0.0426). The accuracy of using these methods is 95.24%, the sensitivity (or normal signal performance) is 100% and the specificity (performance of power disturbance signal) is 94.4%. As a conclusion, SVD and PCA are useful to apply in TFD to extract important feature vectors then MMC can measure the distance metric between those mean vectors. Furthermore, all the features obtained are useful features and can be used for power disturbance classification and recognition with DSP approach as well as to maintain power quality