基于傅里叶神经网络(Fourier neural network,FNN)结构的单输入、输出Hammerstein模型,提出了一种适用于变压器绕组振动系统的非线性建模方法。FNN和自回归滑动平均模型(auto-regressive and moving average model,ARMA)(分别作为模型...基于傅里叶神经网络(Fourier neural network,FNN)结构的单输入、输出Hammerstein模型,提出了一种适用于变压器绕组振动系统的非线性建模方法。FNN和自回归滑动平均模型(auto-regressive and moving average model,ARMA)(分别作为模型中的非线性静态模块和线性动态模块)采用前向更新策略及最速下降法对模型进行训练,确定模型参数。该建模方法在实际110 kV变压器的绕组振动系统的建模及绕组振动波形预测中显示了较高的有效性及准确性。此外,还研究和分析了绕组故障对模型特性的影响机制,提出基于模型延迟阶数的绕组故障特征及其提取方法,并将该方法及其特征量应用于实际110 kV变压器的绕组故障实验中,所得结果表明该特征量可有效反映变压器绕组的机械结构变化。展开更多
Fault detection of an induction motor was carried out using the information of the stator current. After synchronizing the actual data, Fourier and wavelet transformations were adopted in order to obtain the sideband ...Fault detection of an induction motor was carried out using the information of the stator current. After synchronizing the actual data, Fourier and wavelet transformations were adopted in order to obtain the sideband or detail value characteristics under healthy and various faulty operating conditions. The most reliable phase current among the three phase currents was selected using an approach that employs the fuzzy entropy measure. Data were trained with a neural network system, and the fault detection algorithm was verified using the unknown data. Results of the proposed approach based on Fourier and wavelet transformations indicate that the faults can be properly classified into six categories. The training error is 5.3×10-7, and the average test error is 0.103.展开更多
文摘基于傅里叶神经网络(Fourier neural network,FNN)结构的单输入、输出Hammerstein模型,提出了一种适用于变压器绕组振动系统的非线性建模方法。FNN和自回归滑动平均模型(auto-regressive and moving average model,ARMA)(分别作为模型中的非线性静态模块和线性动态模块)采用前向更新策略及最速下降法对模型进行训练,确定模型参数。该建模方法在实际110 kV变压器的绕组振动系统的建模及绕组振动波形预测中显示了较高的有效性及准确性。此外,还研究和分析了绕组故障对模型特性的影响机制,提出基于模型延迟阶数的绕组故障特征及其提取方法,并将该方法及其特征量应用于实际110 kV变压器的绕组故障实验中,所得结果表明该特征量可有效反映变压器绕组的机械结构变化。
基金Project supported by the Second Stage of Brain Korea 21 Projects
文摘Fault detection of an induction motor was carried out using the information of the stator current. After synchronizing the actual data, Fourier and wavelet transformations were adopted in order to obtain the sideband or detail value characteristics under healthy and various faulty operating conditions. The most reliable phase current among the three phase currents was selected using an approach that employs the fuzzy entropy measure. Data were trained with a neural network system, and the fault detection algorithm was verified using the unknown data. Results of the proposed approach based on Fourier and wavelet transformations indicate that the faults can be properly classified into six categories. The training error is 5.3×10-7, and the average test error is 0.103.
