Page 109 - 水利学报2021年第52卷第4期

［ 11］ TRAN Q T，NGUYEN S D，SEO T I . Algorithm for estimating online bearing fault upon the ability to extract
meaningful information from big data of intelligent structures［J］. IEEE Transactions on Industrial Electronics
2019，66（5）：3804-3813 .
［ 12］ 中华人民共和国国家质量监督检验检疫总局，中国国家标准化管理委员会 . GB/T 32584-2016，水力发电厂
和蓄能泵站机组机械振动的评定［S］. 北京：中国标准出版社，2016 .
［ 13］ 吴健健，陈韦晋，章婷婷，等 . 基于智能床垫的心率检测系统［J］. 合肥工业大学学报（自然科学版），2020，
43（3）：330-334，388 .
［ 14］ 卢娜，肖志怀，张广涛，等 . 基于自适应多小波与综合距离评估指数的旋转机械故障特征提取［J］. 振动与
冲击，2014，33（12）：193-199，210 .
［ 15］ 汤 深 伟 ，贾 瑞 玉 . 基 于 改 进 粒 子 群 算 法 的 k 均 值 聚 类 算 法［J］. 计 算 机 工 程 与 应 用 ，2019，55（18）：
140-145 .
［ 16］ 姜伟 . 水电机组混合智能故障诊断与状态趋势预测方法研究［D］. 武汉：华中科技大学，2019 .
［ 17］ XUE X M，ZHOU J Z . A hybrid fault diagnosis approach based on mixed-domain state features for rotating ma⁃
chinery［J］. ISA Transactions，2017，66：284-295 .
［ 18］ 薛小明 . 基于时频分析与特征约简的水电机组故障诊断方法研究［D］. 武汉：华中科技大学，2016 .
［ 19］ 邱锡鹏 . 神经网络与深度学习［EB/OL］.［2019-04-06］. https：/github.com/nndl/nndl.github.io
/
［ 20］ 彭文季，罗兴锜，赵道利 . 基于频谱法与径向基函数网络的水电机组振动故障诊断［J］. 中国电机工程学报，
2006，26（9）：155-158 .

Real-time assessment method of hydropower unit health status
based on unsupervised feature learning

3
1
4
5
HU Xiao ，XIAO Zhihuai 1，2 ，LIU Dong ，WU Daoping ，ZHA Haitao ，LIAO Zhifang 6
（1. School of Power and Mechanical Engineering，Wuhan University，Wuhan 430072，China；
2. Key Laboratory of Hydraulic Machinery Transients，Ministry of Education，Wuhan University，Wuhan 430072，China；
3. State Key Laboratory of Water Resources and Hydropower Engineering Science，Wuhan University，Wuhan 430072，China；
4. State Grid Jiangxi Electric Power Co.，Ltd. Research Institute，Nanchang 330096，China；
5. State Grid Jiangxi Electric Power Co.，Ltd. Zhelin Hydropower Plant，Nanchang 330096，China；
6. Tianjin Key Laboratory of Water Hammer Valve Control Technology Enterprise，Tianjin 300051，China）

Abstract： Real-time assessment of hydropower units’ health status is an important part of condition moni⁃
toring and deterioration warning. Traditional method compares the monitoring value with the limit value of vi⁃
bration， which limits the application range of operation conditions and cannot reflect the individuality of
the hydropower unit. Meanwhile， at present， the known fault types of hydropower units are limited， and
the lack of fault samples brings about difficulties for supervised methods. In this paper， a new method for
constructing vibration degradation indicators using unsupervised feature learning techniques is proposed. The
proposed method makes full use of the large amount of samples in the condition monitoring system of hydro⁃
power units and establishes two degradation indicators：（1） time domain degradation indicator based on fea⁃
ture space reconstruction and singular value decomposition；（2） frequency domain degradation indicator
based on reconstruction error of auto-encoder. Using the proposed degradation indicators can realize re⁃
al-time and quantitative assessment of hydropower units’ health status. A hydropower unit’s axial vibration
waveform dataset has been employed to testify the effectiveness of the proposed degradation indicators. The
results show that the proposed deterioration indicators can effectively reflect the deterioration degree of the
health status of hydropower unit. Moreover，Comparison experiment results indicate that the proposed deterio⁃
ration indicators display the deterioration trend and severity more obviously than the commonly used
time-domain statistical indicators.
Keywords：hydropower units；vibration signal；unsupervised feature learning；singular value decomposition；
auto-encoder；deterioration index
（责任编辑：杨 虹）

— 485 —