基于小波包和BP神经网络的刚性罐道故障诊断

马天兵, 王孝东, 杜菲, 陈南南

(安徽理工大学 机械工程学院, 安徽 淮南 232001)

摘要针对现有刚性罐道故障诊断方法不能消除环境因素影响、接头故障识别率较低等问题,以提高罐道故障种类识别精度为目标,提出了基于小波包和BP神经网络的刚性罐道故障诊断方法。搭建了立井提升系统实验台,模拟台阶突起故障和罐道接头故障这2种典型的罐道故障,采集提升容器振动加速度信号;运用小波包分解对采集的信号进行能量分析并提取故障特征参数,将故障特征参数作为BP神经网络的输入,并选取新的测试样本检测神经网络的诊断效果。测试结果表明,基于小波包分析和BP神经网络的刚性罐道故障诊断方法具有较高的故障识别精度,置信度达到了0.91。

关键词立井提升; 刚性罐道; 故障诊断; 故障种类识别; 小波包; BP神经网络

中图分类号:TD53

文献标志码:A

网络出版地址:http://kns.cnki.net/kcms/detail/32.1627.TP.20180724.0936.001.html

Fault diagnosis of rigid cage guide based on wavelet packet and BP neural network

MA Tianbing, WANG Xiaodong, DU Fei, CHEN Nannan

(School of Mechanical Engineering, Anhui University of Science and Technology, Huainan 232001, China)

Abstract:In view of problems that existing fault diagnosis methods of rigid cage guide could not eliminate influences of environmental factors and low recognition rate of joint faults, a method of fault diagnosis of rigid cage guide based on wavelet packet and BP neural network was proposed in order to improve accuracy of identification of fault types of rigid cage guide. Experimental platform of lifting system of vertical shaft was set up to simulate two typical fault types of rigid cage guide including step protrusion and joint failure, and vibration acceleration signal of lifting vessel was collected. Wavelet packet decomposition was applied to carry out energy analysis and extract fault characteristic parameters. The fault characteristic parameters were taken as input of BP neural network, and a new test sample was selected to detect diagnostic effect of the neural network. The experimental results show that the method has high accuracy of fault identification, and the confidence level reaches to 0.91.

Key words:vertical shaft lifting; rigid cage guide; fault diagnosis; fault type identification; wavelet packet; BP neural network

文章编号1671-251X(2018)08-0076-05

DOI:10.13272/j.issn.1671-251x.2018010051

收稿日期2018-01-18;

修回日期:2018-06-10;

责任编辑:胡娴。

基金项目国家自然科学基金项目(51305003);安徽省博士后基金项目(2017B172);安徽省高校自然科学研究重大项目(KJ2015ZD19);安徽理工大学国家自然基金预研项目(2016yz004)。

作者简介马天兵(1981-),男,安徽合肥人,教授,博士,主要研究方向为智能控制,E-mail:dfmtb@163.com。

引用格式马天兵,王孝东,杜菲,等.基于小波包和BP神经网络的刚性罐道故障诊断[J].工矿自动化,2018,44(8):76-80.

MA Tianbing,WANG Xiaodong,DU Fei,et al.Fault diagnosis of rigid cage guide based on wavelet packet and BP neural network[J].Industry and Mine Automation,2018,44(8):76-80.

0 引言

矿井刚性罐道作为提升容器的导向及限位装置,是立井提升系统的重要组成部分。罐道故障改变了其与提升容器的正常配合,影响乘坐的舒适性和提升容器运行的经济性,严重故障可能导致提升容器脱轨或者卡罐,甚至引起钢丝绳断裂等重大安全事故。

罐道故障诊断方法有几何测距法、专业仪器法和振动加速度测量法等。前两种方法属于静态测试,精确度和经济性较差。振动加速度测量法属于动态测试方法。目前关于罐道故障诊断方法的研究有:Kaczmarczyk S等[1]建立了二自由度的提升容器振动模型并进行了数值计算和仿真;Li Zhangfang等[2]建立了六自由度的提升容器振动模型,提出用小波奇异性来处理振动信号;两者主要从理论上对提升容器进行研究,缺少实验部分。蒋玉强[3]提出采用模糊综合评判法对罐道状态进行评估,但实验信号是在复杂条件下采集的,不能消除环境因素的影响;张淼[4]提出采用支持向量机完成罐道故障识别,但对接头故障的识别率较低。

小波包通过对信号(平稳或非平稳)在不同尺度上的分解与重构,得到原信号在不同频段上的详细分布信息[5-7];BP神经网络被广泛应用于模式识别、函数逼近、分类和数据压缩等,通过对不同状态的信息进行训练而获得某种映射关系[8-9]。提升容器故障处的振动信号是典型的非平稳信号,因此,本文提出了一种基于小波包和BP神经网络的刚性罐道故障诊断方法。通过小波包分解得到不同模式下的故障特征参数,将其作为神经网络输入层,选取不同实验样本对训练好的神经网络进行测试,结果表明该方法有较高的故障识别精度。

1 刚性罐道故障实验系统搭建

为模拟不同罐道故障下提升容器振动响应,搭建如图1所示的立井提升系统,模拟了如图2所示的台阶突起故障和图3所示的罐道接头故障2种故障形式,每种故障采集30组数据。设提升速度为0.18 m/s,模拟不同的罐道故障,采集提升容器的振动加速度信号。信号采集系统包括INV3062T0型四通道数据采集仪、INV98362型三方向传感器和DASP软件。罐道冲击性台阶缺陷引起的高频振动导致的冲击力频率在400 Hz左右,根据采样定理设置采样频率为1 024 Hz。根据文献[3],将传感器贴在提升容器顶部中心。

图1 立井提升系统
Fig.1 Shaft hoist system

图2 台阶突起故障
Fig.2 Step protrusion failure

图3 罐道接头故障
Fig.3 Joint failure of rigid cage guide

虽然在不同故障激励下,提升容器振动响应时域波形会发生变化,但同种故障激励下提升容器振动响应的波形也会发生一定变化,故不能根据响应波形确定故障类型。考虑到提升容器振动信号为典型的非平稳信号,可借助具有全频段分解能力的小波包来进行故障特征识别。

2 刚性罐道故障诊断

2.1 小波包分解与重构

对无故障、台阶突起故障和罐道接头故障3种振动信号进行小波包降噪。不同模式下原始信号与降噪后的信号如图4所示,不同状态下的信号频谱如图5所示。采用db2小波与Shannon熵对降噪后的各种提升容器振动信号进行3层分解与重构[10]。当采样频率为1 024 Hz时,分析频率为512 Hz,在第3层分解为8个频带,各系数对应的频带为0~64,64~128,128~192,192~256,256~320,320~384,384~448,448~512 Hz。对节点系数进行重构,观察各种罐道故障状态下提升容器的响应。经过小波包分解和重构后,信号被分解到8个频段,但各频段总体特征不是十分明显,直接判断故障类型有一定困难,因此,需要借助BP神经网络对各频带的特征进行识别并判断故障类型。

(a) 原始信号(无故障)

(b) 降噪后的信号(无故障)

(c) 原始信号(台阶凸起故障)

(d) 降噪后的信号(台阶凸起故障)

(e) 原始信号(罐道接头故障)

(f) 降噪后的信号(罐道接头故障)

图4 不同模式下的原始信号与经小波包降噪后的信号
Fig.4 Original signal and noise-reduced signal by wavelet packet in different modes

(a) 无故障时的信号频谱

(b) 台阶突起故障时的信号频谱

(c) 罐道接头故障时的信号频谱

图5 不同状态下的信号频谱
Fig.5 Signal spectrums under different conditions

2.2 BP神经网络

BP神经网络是一种多层前馈神经网络,一般由输入层、隐含层和输出层组成,其特点是信号前向传递,误差反向传播。如果输出层得不到期望输出,则转入反向传播,根据预测误差调整网络权值和阈值,从而使BP神经网络预测输出不断逼近期望输出[11]

对采集的信号进行3层小波包分解,得到无故障、台阶凸起故障和罐道接头故障3种模式下8个频带的能量分布,作为BP神经网络的输入,输入节点数为8,输出节点数为3。BP神经网络隐含层的节点数对预测值有较大影响,若节点太少,则训练精度差;若节点太多,则训练时间增加,网络容易出现过拟合现象。经测试,隐含层节点数为5时效果最好,故设隐含层节点数为5。隐含层传递函数采用S型正切函数tansig,输出层传递函数采用purelin函数,训练函数采用Levenberg_Martquardt,设迭代次数为100,学习率为0.1,目标为0.000 04。

将每种模式对应的30组数据作为训练数据。为了简化网络结构,采用归一化处理的特征向量作为输入向量,输出向量(1,0,0)表示罐道无故障,(0,1,0)表示台阶凸起故障,(0,0,1)表示罐道接头故障。部分训练样本见表1。

表1 部分训练样本
Table 1 Partial training samples

数据序号训练样本特征 故障类型输出向量123456(0.2995200,0.3910682,0.0391997,0.2037468,0.0051611,0.0354888,0.0065658,0.0192497)(0.2869903,0.3989111,0.0392834,0.2076018,0.0050152,0.0365675,0.0060391,0.0195915)(0.9496263,0.0374879,0.0086816,0.0018347,0.0011391,0.0000217,0.0003022,0.0009065)(0.9504473,0.0369241,0.0085896,0.0017057,0.0011895,0.0000202,0.0002809,0.0008427)(0.8661933,0.0997131,0.0232027,0.0045855,0.0032303,0.0000542,0.0007552,0.0022656)(0.8642370,0.1011182,0.0234869,0.0047669,0.0031943,0.0000564,0.0007851,0.0023552)无故障(1,0,0)无故障(1,0,0)台阶凸起故障(0,1,0)台阶凸起故障(0,1,0)接头故障(0,0,1)接头故障(0,0,1)

3 测试结果分析

为了验证已建立好的BP神经网络的故障识别效果,选取新的数据作为网络输入。为了评估网络的故障识别精度,定义置信度[12]

式中:yi为网络输出;Yi网络标准输出;n为故障种类,取值为3。

测试样本见表2,测试结果见表3。从表3的测试结果可看出,基于小波包分析和BP神经网络的诊断方法具有较高的故障识别能力,置信度达到0.91。

表2 BP神经网络测试样本
Table 2 Test samples of BP neural network

数据序号测试样本特征 故障类型输出向量123(0.2973925,0.3919778,0.0373398,0.2060911,0.0050185,0.0360045,0.0056676,0.0205081)(0.9513081,0.0362566,0.0084123,0.0017321,0.0011294,0.0000205,0.0002853,0.0008558)(0.8655150,0.1002207,0.0233216,0.0046100,0.0032413,0.0000545,0.0007592,0.0022777)无故障(1,0,0)台阶凸起故障(0,1,0)接头故障(0,0,1)

表3 BP神经网络测试结果
Table 3 Test results of BP neural network

数据序号测试输出理想输出故障类型置信度1(0.9126,0.0847,0.0023)(1,0,0)无故障0.92972(0.0287,0.8545,-0.0033)(0,1,0)台阶凸起故障0.91433(0.0336,-0.0093,0.8726)(0,0,1)接头故障0.9237

4 结语

在刚性罐道不同故障模式下,对提升容器振动加速度时间序列在不同尺度上进行小波包分解与重构,得到不同频段上的能量分布信息。通过基于小波包频带能量分布方法提取不同的故障特征参数,应用BP神经网络诊断方法,对刚性罐道故障类型进行进一步诊断。测试结果表明,基于小波包和BP神经网络的刚性罐道故障诊断方法具有较高的故障识别精度。

参考文献(References):

[1] KACZMARCZYK S,OSTACHOWICZ W. Transient vibration phenomena in deep mine hoisting cables. Part 1:Mathematical model[J]. Journal of Sound and Vibration, 2003,262(2):219-244.

[2] LI Zhanfang,XIAO Xingming ,MA Chi. Study on vibration characteristic of hoist cage and fault diagnosis of steelwork[C]//The 2nd International Conference on Information Science and Engineering, Hangzhou, 2011.

[3] 蒋玉强.立井刚性罐道系统的非线性耦合特性及状态评估研究[D].徐州:中国矿业大学,2011.

[4] 张淼.立井刚性罐道典型故障的模式识别研究[D].徐州:中国矿业大学,2015.

[5] 王浩任,黄亦翔,赵帅,等.基于小波包和拉普拉斯特征值映射的柱塞泵健康评估方法[J].振动与冲击,2017,36(22):45-50.

WANG Haoren,HUANG Yixiang,ZHAO Shuai,et al.Health assessment for a piston pump based on WPD and LE[J].Journal of Vibration and Shock,2017,36(22):45-50.

[6] 吴建萍,姜斌,刘剑慰.基于小波包信息熵和小波神经网络的异步电机故障诊断[J].山东大学学报(工学版),2017,47(5):223-228.

WU Jianping,JIANG Bin,LIU Jianwei.Fault diagnosis of asynchronous motor based on wavelet packet entropy and wavelet neural network[J].Journal of Shandong University(Engineering Science),2017,47(5):223-228.

[7] 罗超,耿蒲龙,曲兵妮,等.基于小波包的矿井供电系统单相接地故障选线方法[J].工矿自动化,2018, 44(2):68-74.

LUO Chao, GENG Pulong, QU Bingni, et al.One-phase grounding fault line selection method based on wavelet packet in mine power supply system[J].Industry and Mine Automation,2018,44(2):68-74.

[8] 谢丽蓉,王晋瑞,穆塔里夫·阿赫迈德,等.基于LVQ-GA-BP神经网络的煤矿瓦斯涌出量预测[J].煤矿安全,2017,48(12):150-152.

XIE Lirong, WANG Jinrui, Mutalif Ahmad, et al. Prediction of mine gas emission based on LVQ-GA-BP neural network[J].Safety in Mines,2017,48(12):150-152.

[9] 梁平,白蕾,龙新峰,等.基于小波包分析及神经网络的汽轮机转子振动故障诊断[J].控制理论与应用,2007,24(6):981-985.

LIANG Ping,BAI Lei,LONG Xinfeng,et al.Turbine rotor vibration faults diagnosis based on wavelet packet analysis and neural network[J].Control Theory & Applications,2007,24(6):981-985.

[10] 赵金宪,金鸿章.基于小波包和神经网络的瓦斯传感器故障诊断[J].传感器与微系统,2010,29(5):80-82.

ZHAO Jinxian,JIN Hongzhang.Gas sensor fault diagnosis based on wavelet packet and neural network[J].Transducer and Microsystem Technologies,2010,29(5):80-82.

[11] 史峰,王小川,郁磊,等. MATLAB神经网络30个案例分析[M].北京:北京航空航天大学出版社,2010.

[12] 陈伟根,邓帮飞.小波包能谱熵与神经网络在断路器故障诊断中的应用[J].重庆大学学报,2008,31(7):744-748.

CHEN Weigen, DENG Bangfei. Application of wavelet packet energy spectrum entropy and neural network to circuit breaker fault diagnosis[J].Journal of Chongqing University,2008,31(7):744-748.