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

doi:10.12096/j.2096-4528.pgt.23114

Abstract

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

CLC Number:

TK 73

Zhong LIU, Zehua ZHOU, Shuyun ZOU, Zhen LIU, Shuaicheng QIAO. Cavitation State Identification Method of Hydraulic Turbine Based on Knowledge Distillation and Convolutional Neural Network[J]. Power Generation Technology, 2025, 46(1): 161-170.

Figures/Tables 17

Fig. 1 CNN structure diagram

Fig. 2 Structure diagram of KD

Fig. 3 KD-CNN model structure diagram

Fig. 4 Flow chart of cavitation state identification

Fig. 5 Schematic diagram of test condition points

Tab. 1 Division of cavitation state of hydraulic turbine

空化状态	空化系数	空化程度
未空化	σ>σ_ini	未发生空化
空化初生	σ最接近σ_ini	未发生明显的空化
空化发展	σ₁<σ<σ_ini	已发生明显的空化
临界空化	σ=σ₁	空化剧烈，效率下降超过1%

Fig. 6 Time domain waveform of cavitation signal of hydraulic turbine

Fig. 7 Spectrum of cavitation signal of hydraulic turbine

Tab. 2 Model parameters of teacher

网络层	核大小	步长	滤波器数量	特征大小
输入层				2 000 $×$ 1
卷积层1	10	1	64	1 991 $×$ 64
池化层1	2	2		995 $×$ 64
卷积层2	10	1	64	986 $×$ 64
池化层2	2	2		493 $×$ 64
卷积层3	10	1	64	484 $×$ 64
池化层3	2	2		242 $×$ 64
Softmax层				4

Tab. 2 Model parameters of teacher

网络层	核大小	步长	滤波器数量	特征大小
输入层				2 000 $×$ 1
卷积层1	10	1	64	1 991 $×$ 64
池化层1	2	2		995 $×$ 64
卷积层2	10	1	64	986 $×$ 64
池化层2	2	2		493 $×$ 64
卷积层3	10	1	64	484 $×$ 64
池化层3	2	2		242 $×$ 64
Softmax层				4

Tab. 3 Model parameters of student

网络层	核大小	步长	滤波器数量	特征大小
输入层				2 000 $×$ 1
卷积层	10	1	2	1 991 $×$ 2
池化层	2	2		995 $×$ 2
Softmax层				4

Tab. 3 Model parameters of student

网络层	核大小	步长	滤波器数量	特征大小
输入层				2 000 $×$ 1
卷积层	10	1	2	1 991 $×$ 2
池化层	2	2		995 $×$ 2
Softmax层				4

Tab. 4 Overall distribution of data

类别	未空化	空化初生	空化发展	临界空化
工况1	3 981	3 981	5 981	3 981
工况2	3 981	3 981	3 981	3 981
工况3	3 981	3 981	3 981	1 981
工况4	3 981	3 981	3 981	3 981

Tab. 5 Cavitation status and the corresponding coding

状态	未空化	空化初生	空化发展	临界空化
编码	[1, 0, 0, 0]	[0, 1, 0, 0]	[0, 0, 1, 0]	[0, 0, 0, 1]

Tab. 6 Data distribution of each working condition

类别	工况1	工况2	工况3	工况4
训练集	14 340	12 740	11 140	12 740
测试集	3 584	3 184	2 784	3 184
总计	17 924	15 924	13 924	15 924

Fig. 8 Teacher model training process curves in condition 2

Fig. 9 Model accuracy comparison curves in condition 2

Fig. 10 Comparison diagram of accuracy of models

Tab. 7 Comparison of time consumption of each model under different working conditions

类别	工况1	工况3	工况4
教师模型	87.36	61.12	80.69
学生模型	1.82	1.62	1.75
KD-CNN模型	1.76	1.56	1.69

References 25

1	LUO X W， JI B， TSUJIMOTO Y ．A review of cavitation in hydraulic machinery[J]．Journal of Hydrodynamics，2016，28(3)：335-358． doi:10.1016/s1001-6058(16)60638-8
2	ESCALER X， EGUSQUIZA E， FARHAT M，et al ．Detection of cavitation in hydraulic turbines[J]．Mechanical Systems and Signal Processing，2006，20(4)：983-1007． doi:10.1016/j.ymssp.2004.08.006
3	RUS T， DULAR M， ŠIROK B，et al ．An investigation of the relationship between acoustic emission，vibration，noise，and cavitation structures on a kaplan turbine[J]．Journal of Fluids Engineering，2007，129(9)：1112-1122． doi:10.1115/1.2754313
4	刘忠，王文豪，邹淑云，等．水轮机空化声发射信号的联合降噪与特征提取[J]．水力发电学报，2022，41(12)：145-152．
	LIU Z， WANG W H， ZOU S Y，et al ．Joint noise reduction and feature extraction of acoustic emission signals for hydraulic turbines under cavitation[J]．Journal of Hydroelectric Engineering，2022，41(12)：145-152．
5	刘君，邓毅，杨延西，等．基于深度学习的空预器转子红外补光图像积灰状态识别[J]．发电技术，2022，43(3)：510-517． doi:10.12096/j.2096-4528.pgt.20122
	LIU J， DEND Y， YANG Y X，et al ．Ash accumulation state identification for infrared compensation images of air preheater rotor based on deep learning method[J]．Power Generation Technology，2022，43(3)：510-517． doi:10.12096/j.2096-4528.pgt.20122
6	武霁阳，李强，陈潜，等．知识图谱框架下基于深度学习的HVDC系统故障辨识[J]．电力系统保护与控制，2023，51(20)：160-169．
	WU J Y， LI Q， CHEN Q，et al ．Fault identification of an HVDC system based on deep learning in the framework of a knowledge graph[J]．Power System Protection and Control，2023，51(20)：160-169．
7	皮志勇，朱益，廖玄，等．基于深度学习的智能变电站通信链路故障定位方法[J]．中国电力，2023，56(7)：136-145．
	PI Z Y， ZHU Y， LIAO X，et al ．Fault location method for communication link in smart substation based on deep learning[J]．Electric Power，2023，56(7)：136-145．
8	封钰，宋佑斌，金晟，等．基于随机森林算法和粗糙集理论的改进型深度学习短期负荷预测模型[J]．发电技术，2023，44(6)：889-895． doi:10.12096/j.2096-4528.pgt.23013
	FENG Y， SONG Y B， JIN S，et al ．Improved deep learning model for forecasting short-term load based on random forest algorithm and rough set theory[J]．Power Generation Technology，2023，44(6)：889-895． doi:10.12096/j.2096-4528.pgt.23013
9	褚雪汝，陈中，吴聪颖，等．基于深度学习的电气二次图纸语义识别方法[J]．浙江电力，2023，42(8)：1-11．
	CHU X R， CHEN Z， WU C Y，et al ．Small target area extraction and semantic recognition method of electrical secondary drawings based on deep learning[J]．Zhejiang Electric Power，2023，42(8)：1-11．
10	钱进宝，刘晓光，蔡玺，等．基于自适应学习率卷积神经网络的新型配电网源网荷储无功协调优化技术[J]．可再生能源，2024，42(2)：267-275．
	QIAN J B， LIU X G， CAI X，et al ．Reactive power coordination optimization technology for source-network-load-storage in new distribution network based on adaptive learning rate convolutional neural network[J]．Renewable Energy Resources，2024，42(2)：267-275．
11	连华，刘明浩．基于多时间尺度卷积神经网络的煤改电用户负荷识别方法研究[J]．电网与清洁能源，2023，39(7)：35-43． doi:10.3969/j.issn.1674-3814.2023.07.005
	LIAN H， LIU M H ．A model of load identification for coal-to-power customers based on multi-time scale convolutional neural networks[J]．Power System and Clean Energy，2023，39(7)：35-43． doi:10.3969/j.issn.1674-3814.2023.07.005
12	娄奇鹤，李荣盛，谭捷，等．基于卷积神经网络的暂稳极限功率计算[J]．中国电力，2024，57(4)：211-219．
	LOU Q H， LI R S， TAN J，et al ．Calculation of transient stability limit based on convolutional neural network[J]．Electric Power，2024，57(4)：211-219．
13	李欣，付豫韬，李新宇，等．基于GAF-CNN的电力系统暂态稳定评估[J]．智慧电力，2023，51(11)：45-52．
	LI X， FU Y T， LI X Y，et al ．Power system transient stability assessment based on GAF-CNN[J]．Smart Power，2023，51(11)：45-52．
14	WEN L， LI X Y， GAO L，et al ．A new convolutional neural network-based data-driven fault diagnosis method[J]．IEEE Transactions on Industrial Electronics，2018，65(7)：5990-5998． doi:10.1109/tie.2017.2774777
15	XU C J， GUAN J J， BAO M，et al ．Pattern recognition based on time-frequency analysis and convolutional neural networks for vibrational events in φ-OTDR[J]．Optical Engineering，2018，57(1)：016103． doi:10.1117/1.oe.57.1.016103
16	贾春阳，曹庆皎，王利英，等．基于卷积神经网络优化的水轮机振动信号识别[J]．噪声与振动控制，2023，43(1)：93-99． doi:10.3969/j.issn.1006-1355.2023.01.016
	JIA C Y， CAO Q J， WANG L Y，et al ．Hydraulic turbine vibration signal recognition based on convolutional neural network optimization[J]．Noise and Vibration Control，2023，43(1)：93-99． doi:10.3969/j.issn.1006-1355.2023.01.016
17	董光德，李道明，陈咏涛，等．基于粒子群优化与卷积神经网络的电能质量扰动分类方法[J]．发电技术，2023，44(1)：136-142． doi:10.12096/j.2096-4528.pgt.22004
	DONG G D， LI D M， CHEN Y T，et al ．Power quality disturbance classification method based on particle swarm optimization and convolutional neural network[J]．Power Generation Technology，2023，44(1)：136-142． doi:10.12096/j.2096-4528.pgt.22004
18	HINTON G， VINYALS O， DEAN J ．Distilling the knowledge in a neural network[EB/OL]．(2015-03-09)[2023-06-05]．．
19	MOHIELDEEN ALABBASY F， ABOHAMAMA A S， ALRAHMAWY M F ．Compressing medical deep neural network models for edge devices using knowledge distillation[J]．Journal of King Saud University：Computer and Information Sciences，2023，35(7)：101616． doi:10.1016/j.jksuci.2023.101616
20	王成林．基于知识蒸馏和深度可分离卷积的轴承故障轻量化诊断[J]．噪声与振动控制，2023，43(3)：139-144． doi:10.3969/j.issn.1006-1355.2023.03.022
	WANG C L ．Lightweight diagnosis for bearing faults based on knowledge distillation and deep separable convolution[J]．Noise and Vibration Control，2023，43(3)：139-144． doi:10.3969/j.issn.1006-1355.2023.03.022
21	ZHANG W F， BISWAS G， ZHAO Q，et al ．Knowledge distilling based model compression and feature learning in fault diagnosis[J]．Applied Soft Computing，2020，88：105958． doi:10.1016/j.asoc.2019.105958
22	JI M Y， PENG G L， LI S J，et al ．A neural network compression method based on knowledge-distillation and parameter quantization for the bearing fault diagnosis[J]．Applied Soft Computing，2022，127：109331． doi:10.1016/j.asoc.2022.109331
23	张哲，秦博宇，高鑫，等．基于CNN-LSTM网络的电网电压稳定紧急控制策略[J]．电力系统自动化，2023，47(11)：60-68． doi:10.7500/AEPS20220712005
	ZHANG Z， QIN B Y， GAO X，et al ．Emergency control strategy of power grid voltage stability based on convolutional neural network and long short-term memory network[J]．Automation of Electric Power Systems，2023，47(11)：60-68． doi:10.7500/AEPS20220712005
24	LIU Y， ZHANG W， WANG J ．Adaptive multi-teacher multi-level knowledge distillation[J]．Neurocomputing，2020，415：106-113． doi:10.1016/j.neucom.2020.07.048
25	刘忠，邹淑云，陈莹，等．混流式水轮机模型空化状态与声发射信号特征关系试验[J]．动力工程学报，2016，36(12)：1017-1022．
	LIU Z， ZOU S Y， CHEN Y，et al ．Experiments on the relationship between cavitation status and acoustic emission signal features for a francis turbine model[J]．Journal of Chinese Society of Power Engineering，2016，36(12)：1017-1022．

[1]	Menghan JIA, Gang LIU, Shijie XU, Shuangying WU. Heterogeneous Image Fusion Algorithm and Its Application in Power Facility Detection Research [J]. Power Generation Technology, 2024, 45(3): 558-565.
[2]	Yu FENG, Youbin SONG, Sheng JIN, Jiahuan FENG, Xuechen SHI, Yongjie YU, Xianchao HUANG. Improved Deep Learning Model for Forecasting Short-Term Load Based on Random Forest Algorithm and Rough Set Theory [J]. Power Generation Technology, 2023, 44(6): 889-895.
[3]	Guangde DONG, Daoming LI, Yongtao CHEN, Xing MA, Ang FU, Gang MU, Bai XIAO. Power Quality Disturbance Classification Method Based on Particle Swarm Optimization and Convolutional Neural Network [J]. Power Generation Technology, 2023, 44(1): 136-142.
[4]	Shangnian CHEN, Luping LI, Shihai ZHANG, Minnan OUYANG, Ang FAN, Xiankui WEN. Research Progress of Vibration Fault Diagnosis Technology for Steam Turbine Generator Sets [J]. Power Generation Technology, 2021, 42(4): 489-499.
[5]	Guodong ZHANG,Haitao PU,Kai LIU. Fault Line Selection Method of Small Current Grounding System Based on Deep Learning [J]. Power Generation Technology, 2019, 40(6): 548-554.
[6]	Xiangjun LI,Gejian XU. Wind Power Prediction Method Based on Long Short-term Memory Neural Network [J]. Power Generation Technology, 2019, 40(5): 426-433.
[7]	Yongli ZHU,Xin SHI,Liuwang WANG. Recent Research Hotspot Introduction on the Application of Artificial Intelligence in Power System [J]. Power Generation Technology, 2018, 39(3): 204-212.