摘要
针对一维卷积神经网络(convolutional neural network,CNN)参数多的特点,提出一种正交试验和粒子群优化算法相结合的参数优化方法,并将其应用于压缩振动信号故障诊断。压缩感知理论突破了奈奎斯特采样定理的限制,为大量振动信号的采集与传输提供一种有效途径。首先利用CNN“端端”特性,建立了基于压缩信号的CNN故障诊断模型。利用正交试验进行参数范围的粗略评价,选择出最优方案。对最优方案中每个参数,利用多目标粒子群优化算法进行细化,得出精确的参数最优取值。选择齿轮箱实测信号和西储大学轴承信号作为研究对象。实验结果表明,经过优化后非劣粒子的输出特征分类明显,CNN诊断率有明显提高,也证明了对压缩信号直接进行故障诊断的可行性。
In view of the characteristics of multi-parameter for the 1-dimensional convolutional neural network(CNN),a parameter optimization method combining orthogonal test and particle swarm optimization algorithm is proposed and applied to the fault diagnosis of compressed vibration signals.Compressive sensing theory breaks through the limitation of Nyquist sampling theorem and provides an effective way for the collection and transmission of a large number of vibration signals.Firstly,a CNN fault diagnosis model based on compressed signal is established by using the“end-to-end”feature of CNN.The orthogonal test is used to make a rough evaluation of the parameter range,and the best scheme is selected.For each parameter in the scheme,the multi-objective particle swarm optimization algorithm is used to refine it,and the optimal value of each parameter is obtained.The measured signal of gear box and bearing signal from Case Western Reserve University are selected as the research objects.The experimental results show that after optimization,the output characteristic classification of non-inferior particles is obvious,and the CNN diagnosis rate is significantly improved.The results also prove the feasibility of direct fault diagnosis for compressed signals.
作者
马云飞
贾希胜
白华军
郭驰名
王双川
MA Yunfei;JIA Xisheng;BAI Huajun;GUO Chiming;WANG Shuangchuan(Department of Equipment Command and Management, Shijiazhuang Campus, Army Engineering University, Shijiazhuang 050003, China)
出处
《系统工程与电子技术》
EI
CSCD
北大核心
2020年第9期1911-1919,共9页
Systems Engineering and Electronics
基金
国家自然科学基金面上项目(71871220)资助课题。
关键词
卷积神经网络
压缩感知
多目标粒子群
正交试验
故障诊断
convolutional neural network(CNN)
compressive sensing
multi-objective particle swarm
orthogonal experiment
fault diagnosis
