基于漏磁负载归一化Lissajous图形分析的变压器绕组故障诊断

Fault Diagnosis Method for Transformer Winding Based on the Load Normalized Lissajous Graphical Analysis of Leakage Magnetic Field

漏磁检测作为变压器故障诊断最具潜力的在线方法之一,由于其漏磁信号受到外部环境和运行条件的影响,实用性还需进一步提升。为了解决这些问题,本文提出了一种基于Lissajous图形与卷积神经网络(CNN)相结合的变压器绕组故障诊断方法。首先,对变压器绕组进行仿真建模,并通过与实验平台测试数据进行一致性验证。然后,设置不同程度和位置的匝间短路故障,收集绕组外部不同位置的漏磁信号。最后,使用本文所提出的LG–CNN方法对绕组故障进行诊断。该方法包括以下3个关键步骤:1)对多工况变压器漏磁信号进行负载归一化;2)将负载归一化后的漏磁信号转换为2维Lissajous漏磁图像;3)使用卷积神经网络对2维漏磁图像进行特征提取并对绕组故障进行诊断。基于漏磁信号的Lissajous图像可以很好地整合各测点漏磁信息之间的关系,针对Lissajous图形随负载变化而变化的问题,本文提出了一种漏磁负载归一化方法,通过实体变压器和高仿真模型实验,验证了所提漏磁负载归一化方法的有效性以及所提检测方法在区分不同程度和位置绕组短路方面的可行性。

Objective Accurately detecting early transformer faults is critical to ensure the stable operation of the power system.Currently,commonly used offline detection methods are limited by maintenance and operational cycles,making real-time monitoring and detecting faults impossible.On-line monitoring methods,such as vibration monitoring and thermal imaging,are constrained by physical structures,and research into detecting weak signal changes that reflect fault characteristics is limited.As a result,these methods cannot accurately detect minor faults inside the transformer.Transformer load variations significantly influence fault classification,leading to reduced classification accuracy.This study proposes a transformer winding fault diagnosis method based on Lissajous graphics and convolutional neural networks(CNN).Methods First,a simulation model consistent with an actual transformer is developed,and the simulation system is utilized to obtain magnetic flux leakage signal data from different measurement points outside the winding under both normal and fault conditions.After processing the collected data,it is randomly divided into a training dataset(validation dataset),and the remaining data is set as a testing dataset.The original signal is then constructed into a high-dimensional space,and parameters such as length,swing angle,area,least square radius,and roundness of the long and short axes in the Lissajous diagram are derived as characteristic parameters that reflect changes in the amplitude and phase angle of the leakage magnetic field.Second,the Lissajous curve is employed to extract the characteristic quantities formed by the leakage magnetic field,and the data is converted into 5×5 grayscale image data.When a fault occurs,the changes in characteristic quantities are extracted as the input for CNN,and artificial intelligence technology is employed to analyze differences in transformer winding fault classifications.This allows for the diagnosis of actual transformer faults using simulation data.Finally,80 simulated data groups for each fault state are used as training samples and 20 waveforms are employed as test samples to verify the effectiveness of the proposed diagnosis method.The results are compared to other models constructed from the original data.Results and Discussions When the transformer model is operated under rated load to produce data,the radial leakage magnetic fields measured at 1 and 5 are plotted as Lissajous diagrams,respectively.The Lissajous number changes with variations in the load,which negatively affects classification accuracy.After normalizing the magnetic flux leakage signal load,the Lissajous diagrams at measurement points 1 and 5 are obtained after load normalization. This normalization eliminates the influence of load-induced changes on the Lissajous number, ensuring that load vari-ations no longer impact the characteristic quantity of the leakage magnetic field. When setting the transformer model to a fault state and using the magnetic flux leakage signal data generated from both normal and fault states, the amplitude and phase angle of the leakage electric field at the measurement point are different. Setting 2-turn, 4-turn, and 8-turn short-circuit faults at the end of the winding of the simulated transformer re-veals that the parameters of the Lissajous diagram change as the severity of the short-circuit fault at the winding end increases. Four parameters are utilized to analyze the Lissajous diagram: long axis (a), short axis (b), slope (k), and eccentricity (e). The differences between the Lissajous diagrams of the normal winding and the fault winding are as follows: the values of parameter a in the 2-turn short-circuit fault, 4-turn (end), 8- turn, and 4-turn (middle) are −0.07%, −0.05%, 0.09%, and 0.04%, respectively. The values of parameter b in the 2-turn short-circuit fault, 4-turn (end), 8-turn, and 4-turn (middle) are 11.5%, 25.42%, −14.01%, and −28.35%, respectively. The values of parameter k in the 2-turn short-circuit fault, 4-turn (end), 8-turn, and 4-turn (middle) are −1.2%, −1.32%, −3.11%, and −1.96%, respectively. The values of parameter e in the 2-turn short-circuit fault, 4-turn (end), 8-turn, and 4-turn (middle) are all 0. The results indicated that the maximum variation occurs in the 4-turn short-circuit fault. When a short-circuit fault occurs, the equivalent circuit undergoes significant changes. However, as the severity of the short-circuit fault increases, the Lissajous diagram does not change linearly as the fault location moves closer to the middle of the winding with increasing fault turns. This indicates that the Lissajous diagram is sensitive to winding faults. The variation in the characteristic quantity during fault occurrence is extracted as the input to CNN by selecting 80 simulation group data samples for each fault state as training samples and 20 waveforms as test samples, converting them into 5×5 grayscale image data. This process is performed on an Intel Core i7-13700k CPU with 32 GB of memory. The deep learning network is constructed using Pytorch, with Python as the programming language. A batch size of 10 data sets, an initial learning rate of 0.001, and 70 training epochs are utilized. The accuracy of the LS-CNN model is 92.61%, that of the LN-SVR model is 95.27%, and that of the LN-LS-CNN model proposed in this study is 97.48%. These results demonstrated that the proposed method is effective for the timely detec-tion and diagnosis of internal transformer winding faults through real-time acquisition of transformer fault data, providing a reference for plan-ning periodic transformer shutdown maintenance. Conclusions The results demonstrated that the proposed method suits transformer magnetic flux leakage detection, even when influenced by load changes. The transformer fault is diagnosed by extracting the characteristic quantity of the leakage magnetic field, improving the utilization of leakage magnetic field information. This method is verified to be effective for the timely detection and diagnosis of internal transformer winding faults through real-time acquisition of transformer fault data and offers a reference for reasonable periodic shutdown maintenance of transformers.