Pattern Recognition for Continuous Partial Discharge Measurements at Power Transformers

在德国使用的许多变压器正接近其预计使用寿命的终点。为了保证尽可能长时间的安全运行，防止故障的发生，必须对绝缘系统的状态进行监测。局部放电测量是绝缘系统状态评估的一种重要的现场测量方法。然而，通过监测系统对pd进行连续测量会产生大量数据，需要对这些数据进行自动处理和分析。为了分析大量的PD数据并进行状态评估，本研究项目将研究模式识别和机器学习方法。在实验室和现场对不同PD源进行测量后，对得到的三维相位分辨图进行分析，提取出鲜明的特征并进行建模。然后通过分类算法将这些特征分配给特定类型的故障。对于气体绝缘开关设备（GIS）和发电机中局部放电的分类，已有一些已知的方法。在本研究项目中，还将研究这些方法是否以及如何用于电力变压器的局部放电诊断。与GIS不同的是，多个局部放电源通常同时激活，并且由于变压器中的外部放电而叠加更多的噪声。为了评估叠加的PD信号，必须首先将它们分离。为此，必须探索适当的方法。由于三维图案可视为具有二维和颜色值的图像，因此除了已知的分类程序外，图像处理方法的使用对于PD的诊断也有待研究。近年来在图像处理领域取得的巨大进展将应用于帕金森病的诊断。因此，PD活动应该被连续和自动地评估。因此，可以在早期阶段预测绝缘的故障，并可以防止变压器的故障。

Many of the transformers used in Germany are approaching the end of their projected service life. To ensure safe operation as long as possible and to prevent failure, the condition of the insulation system has to be monitored. The measurement of partial discharges (PD) is an important on-site measuring method for condition assessment of the insulation system. However, the continuous measurement of the PDs by means of monitoring systems generates large amounts of data, which need to be automatically processed and analyzed.In order to analyze the extensive PD data and enable condition evaluation, methods of pattern recognition and machine learning are to be examined in this research project. After the measurement of different PD sources in the laboratory and on-site, the resulting three-dimensional phase resolved patterns are analyzed, distinctive features are extracted and modeled. These characteristics are then assigned by means of a classification algorithm to a specific class of fault. Some methods are already known for the classification of partial discharges in gas-insulated switchgear (GIS) and generators. In this research project it will also be investigated whether and how these methods can also be used for partial discharge diagnosis in power transformers. Unlike GIS, several PD sources are often active simultaneously and superimposed with more noise by external discharges in transformers. In order to evaluate the superimposed PD signals, they must first be separated. For this, appropriate methods have to be explored. Since three-dimensional patterns can be regarded as images with two dimensions and a color value, in addition to the classification procedures already known, the use of image processing methods are to be examined for the PD diagnosis. The great advances in the field of image processing in recent years will be applied to the PD diagnosis. The PD activity should thereby be evaluated continuously and automatically. Thus the failure of the insulation can be predicted at an early stage and a failure of the transformer can be prevented.