基于卷积神经网络优化的水轮机振动信号识别  

Hydraulic Turbine Vibration Signal Recognition Based on Convolutional Neural Network Optimization

作　　者：贾春阳 曹庆皎 王利英 刘伟 JIA Chunyang;CAO Qingjiao;WANG Liying;LIU Wei(School ofWater Conservancy and Hydropower Engineering,Hebei University of Engineering,Handan 056007,Hebei,China)

机构地区：[1]河北工程大学水利水电学院,河北邯郸056007

出　　处：《噪声与振动控制》2023年第1期93-99,共7页Noise and Vibration Control

基　　金：国家自然科学基金面上资助项目(11972144,12072098);河北省自然科学基金资助项目(E2018402092)。

摘　　要：提出了一种基于蝙蝠算法优化卷积神经网络的水轮机振动信号识别方法。首先对水轮机时域加速度振动信号进行测量、提取和归一化处理,采用蝙蝠算法对卷积神经网络训练过程中的超参数权值和偏置值进行优化,然后对10种不同测点的水轮机振动信号进行实验,针对每个测点的振动信号对水轮机8种不同工况进行区分识别,最后将信号识别过程中各参数对传统卷积神经网络识别结果的影响进行针对性分析。结果表明:所建立的基于蝙蝠算法优化卷积神经网络的识别模型具有良好的稳定性和较高的识别精度,能够准确识别振动信号,识别结果准确率均在94%以上,与传统卷积神经网络对比,信号识别准确率显著提升,最高达到20.78%。同时可以看出,振动数据输入长度、样本尺寸和训练次数对传统卷积神经网络训练效果影响显著。研究结论可为水轮机振动识别、工况识别和故障识别提供理论依据。A vibration signal recognition method based on convolutional neural network optimization with bat algorithm is proposed for hydraulic turbines. Firstly, the turbine vibration acceleration signal is measured, extracted and normalized in time-domain. The bat algorithm is used to optimize the super parameter weights and the offset value in the process of convolutional neural network training. Then, the vibration signals at 10 different measurement points of the turbine are tested, the vibration signal at each measurement point is used to identify the eight different operating conditions of the turbine. Finally, the influence of each parameter in signal recognition process on the recognition result of traditional convolutional neural network is analyzed. Results show that the established recognition model based on the optimization of the convolution of the neural network with bat algorithm has good stability and high accuracy. Its accuracy of vibration signal recognition can be above 94 %. In comparison with the traditional convolution neural network, its signal recognition accuracy rate significantly increases and reaches 20.78 %. Meanwhile, it can be seen that the input length of vibration data,sample size and training times have significant influence on the training effect of traditional convolutional neural network.The conclusion of this work provides a theoretical basis for hydraulic turbine vibration identification, condition identification and fault identification.

关 键 词：振动与波 卷积神经网络 蝙蝠算法 水轮机振动 信号识别 振动参数 

分 类 号：TV734.1[水利工程—水利水电工程]