BP神经网络作业资料.pptx

研究背景及意义 滚动轴承缺陷早期诊断方法的研究，对滚动轴承制造质量的评价，以及设备状态在线监测都有着重要的意义。目前，滚动轴承的故障主要依靠操作人员凭经验进行测试和分析，这给滚动轴承故障诊断的精确实施带来了困难。因此，利用计算机自动进行滚动轴承故障的智能诊断是现代机械工业发展的迫切需要，也是人们追求的目标。自动化智能诊断不仅可以省去操作人员的具体分析工作，还可以快速地在线监测和诊断，并为深人开展机械故障诊断创造了更加有利的条件。人工神经网络是由大量的人工神经元相互连接而成的非线性动力系统，可模仿人脑的智慧进行信息处理。本文提出了一种基于BP神经网络的滚动轴故障自动识别方法，利用该方法可实现滚动轴承故障的智能诊断。滚动轴承特征向量的提取滚动轴承特征向量的提取 对轴承振动信号进行幅域处理常用的指标有：均方根值、峰值、峭度、峰值因子、峭度因子、脉冲因子、裕度因子和波形因子等。其中均方根值和峰值为有量纲参数指标，而峭度、峰值因子、峭度因子、脉冲因子、裕度因子和波形因子为无量纲参数指标。由于有量纲参数指标依赖历史数据并对载荷和转速等的变化比较敏感，而无量纲参数指标基本上不受载荷和转速等因素的影响，无须考虑相对标准值或与以前的数据进行对比，另外，它不受信号绝对水平的影响，即使测量点同以往的略有不同，对参数的计算结果也不会产生明显的影响。因此，选用峭度、峰值因子、峭度因子、脉冲因子、裕度因子及波形因子6个无量纲参数指标来表征滚动轴承运行状态的特征向量。BP神经网络训练样本神经网络训练样本确定网络结构确定网络结构 模型采用3层BP神经网络，输入层为6个节点，对应于峭度、峰值因子、峭度因子、脉冲因子、裕度因子及波形因子6个特征向量，隐层节点数的选取目前尚无理论依据，可根据经验或通过训练学习后，考虑网络的学习次数和识别率综合比较后选定，本文选用隐层节点数为12，输出层节点数为3，对应于轴承的3种故障类型，滚动轴承各种故障类型的期望输出为:正常（0，0，0)，外圈划伤(1，0，0)，内圈划伤(0，1，0)，滚子划伤(0，0，1)。BP神经网络的最终结构为N(6，12，3)。BP网络程序网络程序p=0.0637,0.0410,0.1068,0.1554,0.1894,0;0.0644,0.0396,0.1005,0.1512,0.1908,0.0014;0.0655,0.0403,0.1043,0.1540,0.1866,0.0018;0.0679,0.0413,0.1064,0.1600,0.1848,0.0014;0.0693,0.0424,0.1085,0.1607,0.1950,0.0025;0.0805,0.6150,0.1838,0.2212,0.2940,0.0025;0.0823,0.4501,0.1684,0.2233,0.2912,0.0042;0.0851,0.4995,0.2226,0.2261,0.2895,0.0053;0.0858,0.8152,0.1883,0.2244,0.2874,0.0049;0.0865,0.4918,0.1789,0.2275,0.2919,0.0063;0.1099,0.6311,0.2566,0.3217,0.5142,0.0070;0.1103,0.7200,0.2590,0.3147,0.5212,0.0077;0.1110,0.7749,0.2618,0.3322,0.5072,0.0091;0.1113,0.7452,0.2597,0.3427,0.5317,0.0105;0.1120,0.5754,0.2209,0.3357,0.5247,0.0088;0.2849,0.9209,0.2909,0.4897,0.6682,0.0105;0.2867,0.7263,0.2940,0.5037,0.6297,0.0126;0.2968,0.5978,0.2790,0.5247,0.6472,0.0151;0.3000,1.0000,0.2965,0.1642,0.6962,0.0165;0.3035,0.8089,0.2986,0.5072,0.6682,0.0144;t=0,0,0;0,0,0;0,0,0;0,0,0;0,0,0;1,0,0;1,0,0;1,0,0;1,0,0;1,0,0;0,1,0;0,1,0;0,1,0;0,1,0;0,1,0;0,0,1;0,0,1;0,0,1;0,0,1;0,0,1;p=p;t=t;net=newff(minmax(p),12,3,tansig,logsig,trainlm);net.trainparam.epochs=100;net.trainparam.goal=0.001;net.trainParam.lr=0.05;net=train(net,p,t);p_test=0.0791,0.0497,0.1247,0.1901,0.2291,0;0.0837,0.0526,0.1367,0.2034,0.2486,0.0041;0.0965,0.5911,0.2179,0.2552,0.3302,0.0054;0.1036,0.7063,0.2075,0.2862,0.3869,0.0075;0.1321,1.0000,0.3099,0.3931,0.6417,0.0095;0.1616,0.6810,0.2614,0.5215,0.7038,0.0133;0.3451,0.3451,0.2887,0.6210,0.7659,0.0137;0.3567,0.3567,0.3509,0.5671,0.9068,0.0203;p_test=p_test;y=(sim(net,p_test)BP网络的训练网络的训练检验网络检验网络轴承状态正常外圈划伤内圈划伤滚子划伤峭度3.123.233.543.714.405.119.549.82峭度因子2.412.1415.4818.2625.3517.659.549.82峰值因子4.224.516.476.228.697.528.189.68脉冲因子5.806.127.378.1210.713.816.214.9裕度因子6.747.219.1810.5516.718.219.723.1波形因子1.211.311.341.391.441.531.541.7理想输出（0,0,0）（0,0,0）（1,0,0）（1,0,0）（0,1,0）（0,1,0）（0,0,1）（0,0,1）实际输出（0.0076,0.0375,0.0001）(0.0101,0.0099,0.0002)(0.9397,0.0485,0.0044)(0.8526,0.1319,0.0104)(0.0002,0.9618,0.1582)(0.0001,0.9901,0.3373)(0.0002,0.0668,0.9525)(0.0000,0.4992,0.9759)用训练好的网络对检验样本进行识别,神经网络对滚动轴承检验样本进行诊断的输出结果如下表所示，可以看出，神经网络能根据所测的数据准确地判断轴承的故障类型，说明神经网络具有很强的识别能力，诊断效果良好。结束语结束语 轴承是机械系统中重要的支承部件,轴承故障诊断是一个非常复杂的问题,轴承受安装位置、运行环境等因素的影响,其故障与征兆间的关系不是很明确,而BP神经网络模型具有较强的自学习、自适应、联想记忆及非线性模式识别的能力,特别适用多故障、多征兆等复杂模式的识别。本文通过对轴承的振动信号进行处理计算,提取有用的无量纲特征参数作为神经网络的输入,实验结果表明,利用BP神经网络对滚动轴承故障模式进行识别是可靠和有效的。