Application of Wavelet Neural Network to Fault Diagnosis of Power System

Jian Hu, Fanjun Hu
2015 Proceedings of the 2015 International Conference on Automation, Mechanical Control and Computational Engineering   unpublished
Based on analyzing the incompleteness and uncertainty of information existing in power system fault diagnosis, a new fault diagnosis method based on wavelet transform and neural network is proposed. The wavelet transform is used to pre-process data and extract feature vectors. The neural network is used to identify fault types. Diagnostic results of instance proved the effectiveness and superiority of the proposed method. Introduction With the rapid development of scientific technology, the
more » ... e and structure of power system continue to expand and become complicated. In the process of power system operation, natural and man-made interference often occurs and the failure is difficult to avoid. Therefore, adopting effective method to diagnose the fault of power system accurately, finding out the fault components, reducing manpower, material resources and economic losses, appear particularly important. Now, with the further development of artificial intelligence, especially machine learning, data mining, etc, many theories and methods are offered to diagnose the fault. Such as expert system, optimization method, fuzzy sets theory etc. Although these theoretical researches have scored some achievements, there are still certain limitations. For example, Fourier transform has been playing an important role in data processing, but Fourier transform has some defects, on the one hand, it can only analyze stationary signals, it cannot characterize sharp-variation signals that occur during faults diagnosing. On the other hand, Fourier transform cannot localize the singularities that always symbolize some sudden faults, and its frequency and time resolutions contradict each other. As a powerful tool of signal analysis, wavelet transform has good localization properties in time and frequency domain, focus to any details of the analysis object with taking fine time or frequency step length of high frequency, express any changes existing in the object, so as to get accurate feature separation results from the measurement data with bad SNR[1-5]. By using wavelet transform to separate the feature, the key process lies in the determination of optimal decomposition levels. On the one hand, we want to separate the feature components as far as possible, on the other hand, keep the fixed errors and true value apart from the separated feature. The current methods need either manual setting threshold control or results testing with extracted trend by wavelet transform, which increase the difficulty of the application of separation methods and raise the risk of error introduced. In accordance with the above case, the paper proposes a new method which approximates the feature with detail components of the wavelet decomposition, determines the optimal decomposition level on the frequency intervals between the feature and other components, then gets the feature directly. The method avoids the indirect error with modeling and indirect methods [6, 7] . The uncertainty of power system operation, the diversity, complexity and associated leveloriented of gathering information, cause detection randomness and uncertainty. Wavelet is a new developing signal processing means. It is localized both in time and frequency domains. So it is possible to characterize the local singularities based on the coefficients in a wavelet orthonormal basis expansion. Combining model theory and statistical knowledge, wavelet provides a method to describe causal relationship between variables. Using probability theory to handle the uncertainty between different knowledge for conditions related, so it thus becomes one of the models in the field of uncertain knowledge representation and reasoning. Applying the neural network to power system fault diagnosis, can solve incomplete and uncertainty. Using the neural network structure
doi:10.2991/amcce-15.2015.409 fatcat:pdhh6rxu3fc5vcyebsfpmz2qze