A Generalized Empirical Wavelet Transform for Classification of Power Quality Disturbances Karthik Thirumala, Student Member, IEEE, Amod C. Umarikar and Trapti Jain, Senior Member, IEEE Discipline of Electrical Engineering Indian Institute of Technology Indore Indore, India Email: {phd1301102004, umarikar, traptij}@iiti.ac.in Abstract—This paper proposes a generalized empirical wavelet transform (GEWT) for the recognition of single and combined power quality (PQ) disturbances. The FFT based frequency estimation is adaptive, requires no prior information and is also capable to diagnose all the PQ disturbances. The im- proved spectral segmentation followed by an adaptive filter design accurately extracts the fundamental frequency component, thereby enabling the extraction of informative features. Thus, the proposed approach combines the GEWT and a simple rule based decision tree (DT) for accurate recognition of most significant PQ disturbances. The DT classifier with five features extracted from the GEWT is refined and finally tested on 1200 simulated as well as three real disturbance signals. The proposed scheme is found to be computationally efficient and performs satisfactorily with a good classification accuracy. Index Terms—Generalized Empirical Wavelet Transform (GEWT), Decision Tree, Power Quality (PQ), PQ Disturbances. I. I NTRODUCTION Quality of electric power is usually deteriorated due to the occurrence of disturbances such as voltage sag, swell, interrup- tion, transients, harmonics and flicker [1]. These disturbances result in mal-operation of circuit breaker, failure of end-user equipment and degrade the performance of transmission sys- tem equipment. Identification of the disturbances and its source is preliminary to deploy a mitigation device for improving the PQ. Most of the traditional power analyzers do not provide sufficient temporal information of the disturbances. Hence, the monitoring equipment should be capable to recognize and classify the PQ disturbances [2]. The signal processing techniques frequently employed for feature extraction include fast Fourier transform (FFT), Wavelet transform (WT), S-transform (ST), empirical mode decomposition (EMD) and Kalman filter [3]–[6]. The features are then utilized as input to a classifier such as expert system, rule base, fuzzy logic (FL), artificial neural network (ANN) and support vector machine (SVM) [7]–[11]. Various approaches have been proposed with different combinations of signal processing techniques and classifiers to enhance the accuracy of classification [12]. The approximation coefficients of the WT are not unique and contain information of a band of frequencies including the fundamental frequency. Similarly, the ST and EMD suffer from mode mixing and the features extracted from these components may not reflect the disturbances acceptably. It can also be inferred that for an accurate recognition of disturbances, signal decomposition with predefined filters necessities unique novel features or a complex classifier. Nevertheless, a linear classifier with basic features is adequate to identify the significant PQ disturbances if the filtering approach is adaptive and accurate. The rule based decision tree (DT) classifier requires less computational resources and is flexible. The empirical wavelet transform (EWT) proposed in [13] decomposes a non-stationary signal by designing a bank of filters according to the frequency information. However, it requires prior information of number of frequencies present in the signal which is impractical for continuous monitoring. This has been overcome in [14], where a EWT based approach, proposed for PQ application, uses minimum magnitude (dM ) and minimum frequency separation threshold (dF ) for esti- mation of actual frequencies. However, the fixed frequency threshold is not suitable to analyze all sorts of PQ signals and may result in inaccurate extraction of fundamental frequency component in case of disturbances. Therefore, this paper pro- poses a generalized EWT (GEWT) with an adaptive frequency threshold for an improved spectral segmentation. The paper aims to recognize the significant PQ disturbances with a simple DT by using features extracted from the GEWT. The developed GEWT first computes the dF threshold re- quired for the FFT based frequency estimation. This novel dF , is estimated adaptively to separate the fundamental fre- quency components from its neighbouring frequencies. The information about the existence of frequencies in vicinity of fundamental is obtained by a primary investigation of the signal amplitude variation, as explained in section II. The improved spectral segmentation followed by adaptive filtering extracts the frequency components with minimal overlapping; thereby the features can represent the disturbances. The major contributions of this paper are new GEWT algorithm for fast and accurate decomposition of disturbance signals and an efficient, simple DT classifier to identify them. II. PROPOSED METHODOLOGY A. Generalized Empirical Wavelet Transform The EWT approach proposed in [14] decomposes the signal into its individual components perfectly but may result in indefinite components in case of disturbance signals. This is the consequence of fixing the minimum frequency separation 978-1-4673-8848-1/16/$31.00 c 2016 IEEE