由于滚动轴承的工作环境复杂,所采集的信号中通常含有大量噪声,噪声的存在会影响故障诊断的结果.为了提高噪声数据的诊断精度,采用改进的小波阈值函数结合栈式自编码器(stacked auto-encoder,SAE)对强噪声环境下的轴承数据进行故障诊断.首先通过改进阈值函数对噪声数据进行去噪,其次用小波包变换提取去噪数据的小波包能量,最后通过SAE得到故障的分类结果.通过在凯斯西储大学的轴承数据集上的实验表明,该模型能够在强噪声背景下得到较为准确的分类结果.

卷积神经网络(Convolution Neural Network,CNN)是一种常用的智能故障诊断方法.针对卷积神经网络结构中参数较多,训练时间长,并且sigmoid和ReLU激活函数运用带来梯度消失和均值偏移.基于以上问题提出了一种改进非线性映射函数的卷积神经网络模型.把振动信号转换成二维振动图像表示其故障纹理,在卷积层中加入了两层残差神经元.再运用该方法对凯斯西储大学和帕德尔伯恩大学轴承数据集进行分析,分别达到预测精度99.18％和100％.

针对长短时记忆网络(Long Short Term Memory,LSTM)处理大数据集时运行时间长、存在维数灾难的问题,提出基于能量熵和CL-LSTM(Long Short Term Memory Network with Center Loss)的智能故障诊断模型.利用自适应白噪声的完整集合经验模态分解对原始信号进行分解;结合相关系数筛选IMF分量并计算其能量熵作为新样本输入到LSTM中,增强了样本间的差异性,减小了数据维度.将中心损失引入Softmax损失中,使类内距离更小,进一步提高分类精度.利用西储大学轴承数据集进行实验,验证了所提方法在识别滚动轴承故障状态时准确率高、稳定性好.

在实际应用中,滚动轴承大多时候都是在正常状态下工作,因此收集到的故障数据较少,这就会产生数据不均衡的问题.这种数据不均衡问题极大地影响着模型的拟合和泛化能力,导致模型产生过拟合情况,而往往忽视对小类别样本的学习.尤其当故障样本数极少时,此问题更突出.针对这个问题,提出一种基于改进交叉熵损失函数的深度自编码器的诊断模型,首先提取振动数据的小波包能量,其次将小波包能量输入到深度自编码器中,最后通过SoftMax分类器得到诊断结果.改进的加权损失函数可以根据各类别样本的数量调整权重系数,样本数量越少,系数越大,使得模型在训练时更专注于数量较少的样本.通过在凯斯西储大学及西安交通大学的轴承数据集上的两个实验表明,加权损失函数可以提高极端不均衡数据的诊断精度.

Vibration signal is the main measurement signal of mechanical component fault diagnosis, and the presence of noise affects the feature extraction of the signal and the final fault diagnosis, so the test signal in practice needs to be denoised. The use of wavelet transform does not filter out the noise in the signal very well. This is because the hard threshold function is not continuous and some useful information is filtered out. There is a deviation in the soft threshold function and the noise in the signal cannot be completely filtered out. And the traditional threshold is fixed. In order to solve the problem of threshold function and threshold, this paper proposes a new threshold function, and uses artificial fish swarm algorithm to get the optimal threshold. Finally, the superiority and practicability of the method are verified by the unsteady test signal and the bearing dataset of Case Western Reserve University. From the final noise reduction result, the method can achieve better performance in noise reduction than other existing methods. The signal-to-noise ratio obtained by this method is 13%~16% higher than other methods. The root mean square error is 10%~41% lower than other methods.

Research on the method of measuring the square wave voltage of oscilloscope calibrator based on digital voltmeter