发电技术 ›› 2025, Vol. 46 ›› Issue (1): 161-170.DOI: 10.12096/j.2096-4528.pgt.23114

• 发电及环境保护 • 上一篇    

基于知识蒸馏-卷积神经网络的水轮机空化状态识别方法

刘忠, 周泽华, 邹淑云, 刘圳, 乔帅程   

  1. 长沙理工大学能源与动力工程学院,湖南省 长沙市 410114
  • 收稿日期:2023-12-13 修回日期:2024-03-02 出版日期:2025-02-28 发布日期:2025-02-27
  • 作者简介:刘忠(1978),男,博士,教授,研究方向为水力机械状态监测与故障诊断,drliuzhong@csust.edu.cn
    周泽华(1999),男,硕士研究生,研究方向为水力机械状态监测与故障诊断,93748226@qq.com
    邹淑云(1979),女,硕士,实验师,研究方向为流体机械内部流动与优化设计,cslgdxzsy@csust.edu.cn;
    刘圳(1998),男,硕士研究生,研究方向为水力机械状态监测与故障诊断,1203235718@qq.com
    乔帅程(1998),男,硕士研究生,研究方向为水力机械状态监测与故障诊断,1771862910@qq.com
  • 基金资助:
    国家自然科学基金项目(52079011);湖南省自然科学基金项目(2023JJ30032);湖南省研究生科研创新项目(CX20220927)

Cavitation State Identification Method of Hydraulic Turbine Based on Knowledge Distillation and Convolutional Neural Network

Zhong LIU, Zehua ZHOU, Shuyun ZOU, Zhen LIU, Shuaicheng QIAO   

  1. School of Energy and Power Engineering, Changsha University of Science and Technology, Changsha 410114, Hunan Province, China
  • Received:2023-12-13 Revised:2024-03-02 Published:2025-02-28 Online:2025-02-27
  • Supported by:
    National Natural Science Foundation of China(52079011);Natural Science Foundation of Hunan Province(2023JJ30032);Hunan Graduate Research Innovation Project(CX20220927)

摘要:

目的 当水轮机发生空化故障时,机组效率下降、部件侵蚀加速,严重时甚至引发安全事故。因此,准确且快速识别水轮机空化状态,对水电站高效、安全运行至关重要。针对目前复杂卷积神经网络(convolutional neural network,CNN)模型存在的识别速度慢与简单CNN模型存在的识别准确率低等问题,提出一种基于知识蒸馏(knowledge distillation,KD)-CNN的水轮机空化状态识别方法。 方法 首先,引入知识蒸馏理论中教师模型与学生模型相互作用机理,定义3层CNN网络作为教师模型,定义单层CNN网络作为学生模型;然后,利用试验获取的空化声发射信号数据对教师模型进行训练;最后,将代表空化状态类型的数据标签替换成教师模型的输出,通过学生模型对替换标签后的新数据集进行学习,使交叉熵达到最小值。训练完成后的模型即为KD-CNN模型,利用该模型对各工况数据进行空化状态识别试验。 结果 KD-CNN模型在2 s内即可完成水轮机空化状态识别,且各工况的识别准确率均高于97%。 结论 KD-CNN模型结构简单,同时具有学生模型的识别速度与教师模型的识别准确率,为水轮机空化实时监测提供了新思路。

关键词: 水轮机空化, 卷积神经网络(CNN), 知识蒸馏(KD), 声发射信号, 深度学习, 状态识别

Abstract:

Objectives When cavitation faults occur in hydraulic turbines, the efficiency of the unit decreases, component erosion accelerates, and in severe cases, safety accidents may even be triggered. Therefore, accurately and rapidly identifying the cavitation state of hydraulic turbines is crucial for the efficient and safe operation of hydropower stations. Aiming at the problems of slow recognition speed in complex convolutional neural network (CNN) models and low recognition accuracy in simple CNN models, a turbine cavitation state recognition method based on knowledge distillation and convolutional neural network (KD-CNN) is proposed. Methods Firstly, the interaction mechanism of teacher model and student model in knowledge distillation theory are introduced, and the teacher model is defined as three-layer CNN network, while the student model is defined as single-layer CNN network. Then, the cavitation acoustic emission signal data obtained from the experiment are used to train the teacher model. Finally, the data labels representing the types of cavitation state are replaced with the output of the teacher model, and new dataset are learned by the student model to minimize the cross entropy. The trained model is the KD-CNN model, which is used for cavitation state recognition experiments on various conditions. Results The KD-CNN model can complete the identification of the cavitation state of the turbine within 2 s, and the recognition accuracy of each working condition is higher than 97%. Conclusions The KD-CNN model has a simple structure, the recognition speed of the student model, and the recognition accuracy of the teacher model. It can provide a theoretical basis for real-time monitoring of cavitation in hydraulic turbines.

Key words: hydraulic turbine cavitation, convolutional neural network (CNN), knowledge distillation (KD), acoustic emission signal, deep learning, state recognition

中图分类号: