Frequency-Domain Dataset for Inter-Turn Short-Circuit (ITSC) Fault Detection in EDF Low-Speed Synchronous Generators

Content This dataset comprises measurements collected from six low-speed synchronous generators located in different hydroelectric power plants operated by EDF, denoted A through F. For each machine, 15 operating points were recorded, resulting in a total of 90 instances acquisitions across the six generators. Each instance is described by 16 columns: 14 frequency-domain features (ranging from 3.3 Hz to 46.7 Hz), a binary label y (where 0 indicates the absence of inter-turn short circuit (ITSC) and 1 indicates its presence), and the machine identifier (A–F). Machine characteristics The main characteristics of the machines are reported below. Characteristic Machines A, B, C Machines D, E, F Rated active power (MW), $P_n$ 55 65 Rated reactive power (MVAR), $Q_n$ 35 50 Rated stator current (A) 3000 3700 Number of poles, $p$ 30 30 Rotational speed (rpm) 200 200 Supply voltage (V) 10000 10000 Frequency (Hz), $m$ 50 50 Data acquisition The dataset was collected by measuring the stray magnetic field around each generator using a tri-axial flux sensor. This sensor consists of three orthogonal coils aligned along the normal (NOR), axial (AXI), and tangential (TAN) directions, and is attached to the external frame of the machine using magnetic supports. The sensor is directly connected to the acquisition system, without additional signal conditioning. In this study, only the tangential (TAN) component is considered. This component, oriented perpendicular to the machine’s longitudinal axis, is less sensitive to axial magnetic effects (e.g., from winding ends) and provides an indirect measurement of the radial air-gap flux, although attenuated by the stator structure. Measurements were conducted over 15 operating points per machine, defined by combinations of three active power levels ($10\%$, $50\%$, and $100\%$ of the nominal power $P_n$) and five reactive power levels ($-100\%$, $-50\%$, $0\%$, $+50\%$, and $+100\%$ of the nominal reactive power $Q_n$). For each operating condition, the sensor signals were recorded using a National Instruments PXI system controlled via a LabVIEW interface for raw data acquisition. Each recording spans 5 seconds at a sampling rate of 20 kHz, producing a univariate time series of length $T = 100 000$. These raw time series are not available in this dataset. Data transformation The raw time-series signals were converted into frequency-domain features using the discrete Fourier transform (DFT). Only harmonics corresponding to integer multiples of the fundamental electrical frequency $m$ (50 Hz), scaled by the number of pole pairs $\frac{p}{2}$ (15 in this case), and falling within the lower sideband (i.e., below the fundamental frequency) were retained for analysis. The feature set is constructed by normalizing the amplitudes of these harmonics with respect to the fundamental component.This procedure results in a feature vector of 14 normalized frequency components for each signal. Acknowledgments This work has benefited from the support of EDF and the National Research Agency under France 2030, MAIA Project ANR-22-EXES-0009.数据集内容 本数据集包含法国电力公司（EDF）运营的不同水电站中6台低速同步发电机的采集数据，发电机编号为A至F。 针对每台发电机，共记录15个运行工况点，6台发电机总计得到90个采集样本。 每个样本由16个字段构成：14个频域特征（频率范围3.3 Hz至46.7 Hz）、一个二元标签$y$（其中0表示不存在匝间短路（inter-turn short circuit, ITSC），1表示存在该故障），以及发电机编号（A至F）。 ## 发电机参数特性 各发电机的主要参数如下： 1. 额定有功功率（MW），记为$P_n$：机组A、B、C为55，机组D、E、F为65； 2. 额定无功功率（MVAR），记为$Q_n$：机组A、B、C为35，机组D、E、F为50； 3. 额定定子电流（A）：机组A、B、C为3000，机组D、E、F为3700； 4. 极数$p$：机组A、B、C和D、E、F均为30； 5. 转速（转/分钟）：机组A、B、C和D、E、F均为200； 6. 供电电压（V）：机组A、B、C和D、E、F均为10000； 7. 供电频率（Hz），记为$m$：机组A、B、C和D、E、F均为50。 ## 数据采集方案 本数据集通过三轴磁通传感器测量每台发电机周围的杂散磁场进行采集。该传感器包含三个沿法向（normal, NOR）、轴向（axial, AXI）和切向（tangential, TAN）方向排布的正交线圈，通过磁性支架固定在发电机外部机座上。传感器直接与采集系统相连，无需额外的信号调理电路。 本研究仅采用切向（TAN）分量。该分量垂直于发电机纵轴方向，对轴向磁场效应（例如绕组端部产生的磁场）敏感度较低，可间接反映径向气隙磁通，尽管该磁通会被定子结构衰减。 每台发电机的测量覆盖15个运行工况点，这些工况点由3个有功功率档位（额定有功功率$P_n$的10%、50%和100%）与5个无功功率档位（额定无功功率$Q_n$的-100%、-50%、0%、+50%和+100%）组合而成。 针对每个运行工况，传感器信号通过美国国家仪器（National Instruments）PXI系统进行记录，该系统由LabVIEW界面控制以完成原始数据采集。每段记录时长为5秒，采样率为20 kHz，生成长度$T=100000$的单变量时间序列。本数据集未提供这些原始时间序列数据。 ## 数据转换流程 原始时间序列信号通过离散傅里叶变换（discrete Fourier transform, DFT）转换为频域特征。本研究仅保留与基频$m$（50 Hz）的整数倍谐波相关的特征，且该谐波需经过极对数$frac{p}{2}$（本案例中为15）缩放后处于下边带（即低于基频）范围内，方可用于分析。 特征集通过将上述谐波的幅值相对于基波分量进行归一化构建而成。该处理流程为每个信号生成包含14个归一化频域分量的特征向量。 ## 致谢 本研究得到了法国电力公司（EDF）以及法国‘未来投资计划（France 2030）’框架下的国家研究署（National Research Agency）支持，具体项目为MAIA项目ANR-22-EXES-0009。