隨著電力系統電壓等級的提升，局部放電可能造成高壓電力設備之絕緣劣化。近年來，局部放電的檢測和圖譜辨識已成爲預防性設備故障診斷的最新發展趨勢；因此，若能結合局部放電檢測與訊號分析，掌握電力設備的絕緣狀態，當能及時避免電力設備的無預警停機，增加供電品質的可靠度。
本文旨在應用類神經網路於高壓馬達定子絕緣線棒局部放電圖譜之辨識。首先，使用高頻電流感測器量取被試物的局部放電訊號，並將接收的局部放電訊號轉換為能量圖譜，以抑制其外在雜訊。其次，使用分形理論萃取圖譜中的分形維數與間隙度二種特徵，以降低維數利於後續運算。最後，利用類神經網路進行局部放電圖譜辨識，即由擷取出的特徵向量當輸入，決定最佳類神經網路架構，以獲取最佳之辨識能力。本研究以常見的高壓馬達定子故障型態，試製四種有絕緣瑕疵的定子試驗模型，並與健康馬達模型做相互對照。實驗結果顯示，本文提出基於類神經網路之定子故障診斷系統，使用共軛梯度演算法在隱藏層神經元數目為20的情況下，可獲得高達90 %之辨識率。

As the power system voltage class is upgraded, the partial discharge may cause insulation deterioration of high voltage power equipments. In recent years, the detection of partial discharge and pattern recognition have become the latest development trend of preventive equipment fault diagnosis. Therefore, if the partial discharge detection can be combined with signal analysis to master the state of insulation of power equipments, the shutdown of power equipments without early warning is supposed to be prevented in time, so as to enhance the reliability of power quality.
This thesis aims to use artificial neural network for recognition of partial discharge pattern of high-voltage motor stator insulation coil bar. First, the partial discharge signal of the test object is measured by using high-frequency current sensor, and the received partial discharge signal is changed into energy spectrum, so as to suppress the external noise. Secondly, the fractal dimension and lacunarity features are extracted from the spectrum by using fractal theory, so as to reduce the dimension for subsequent operation. Finally, the artificial neural network is used for partial discharge pattern recognition. The extracted features vector is used as input and the optimum Artificial Neural Network is determined, so as to obtain the optimum recognition capability. This study uses common high voltage motor stator fault types to experimentally produce four kinds of stator test models with insulation defect, which are compared with healthy motor model. The experimental results show that the artificial neural network-based stator fault diagnosis system proposed in thesis has a recognition rate as high as 90 % when the conjugate gradient algorithm is used and there are 20 neurons on hidden layer.

[1]“High-voltage test techniques – Partial discharge measurements,” CNS 15175, 2008.
[2]“High voltage test techniques – Partial discharge measurement,” IEC 60270, 2001.
[3]Jian Li, Tianyan Jiang, and Robert F. Harrison, “Recognition of ultra high frequency partial discharge signals using multi-scale features,” IEEE Transactions on Dielectrics and Electrical Insulation, Vol. 19, No. 4, pp. 1412-1420, 2012.
[4]Liwei Hao and Paul Lewin, “Partial discharge source discrimination using a support vector machine,” Rotating electrical machines – Part 27-2: On-line partial discharge measurements on the stator winding insulation of rotating electrical machines,” IEEE Transactions on Dielectrics and Electrical Insulation, Vol. 17, No. 1, pp. 189-197, 2010.