| Summary: | 碩士 === 元智大學 === 電機工程學系 === 106 === With the vigorous development of new energy generation and green intelligent electricity utilization, more and more power electronic converters are applied to smart power grid and become the key to energy transformation. As the core of energy conversion, the failure of power electronic converters may cause the whole system to fail or even paralyze, and result in serious losses. Therefore, it is very important to study the fault diagnosis method of converter circuit in depth; predict the fault risk with the aid of intelligent diagnosis method; realize intelligent, fast and accurate diagnosis of the fault type of power electronic converter; locate fault components; replace fault components in advance; and ensure the reliability and stability of the system operation.
Based on the fault characteristics of power electronic converter circuit, the fault diagnosis and location of converter circuit are studied in this thesis. By analyzing the overall output signal of the converter circuit, the methods of fault feature extraction, fault diagnosis and location are studied. The main research is as follows:
(1) A fault feature extraction method of converter circuit based on fast Fourier transform is studied. First, the fault features are extracted by fast Fourier transform, and fault feature vector is formed. On this basis, using principal component analysis method to further screen the fault feature vectors, extracting the principal components of the feature vector that can represent the operating state of the converter circuit, reduce the redundant fault features, and reduce the data dimension.
(2)Proposed a kind of intelligent fault diagnosis methods with high diagnostic accuracy and high speed—based on cerebellar model neural network (CMNN) fault diagnosis method of the converter circuit. By using the back-propagation algorithm to update neural network parameters, the CMNN can make the diagnosis error converge quickly. After off-line training, the CMNN diagnosis unit can accurately identify the type of circuit fault and the specific fault location through fault features, then it can obtain accurate diagnosis results.
(3) Aiming at the disadvantages of poor initial parameter selection, which has great influence on the learning ability of neural network diagnostics, a global optimization search method, Genetic Algorithm, is used to optimize the initial weights of neural network to obtain better initial parameters. Through further training, the neural network internal parameters (weights, deviations) are approached faster and more accurate toward the value that produces the smallest error.
(4) Taking the dual-buck bidirectional DC-AC converter circuit as an example, the fault diagnosis simulation of the converter circuit is carried out. The equivalent model of components that is prone to parametric faults is established, and set different levels of fault types. Under different working conditions, the performance of the fault diagnosis unit based on the CMNN for the parametric fault diagnosis of dual-buck bidirectional DC-AC converter circuit is simulated. The diagnostic results are compared with the that using a traditional BP neural network diagnostic unit. The comparison results show that the proposed fault diagnosis method can identify fault types more quickly and has higher diagnostic accuracy.
|
|---|
