一种深度卷积自编码网络及其在 滚动轴承故障诊断中的应用
A Deep Convolutional Auto？？Encoding Neural Network and Its Application in Bearing Fault Diagnosis

为了解决卷积神经网络权值往往只能随机初始化的问题,提出了一种卷积自编码器。以卷积池化过后的特征为权值,对反卷积核进行叠加,叠加步长为池化时的长度,将信号重构回原信号空间。以原信号与重构信号的差值最小为目标,对卷积核和反卷积核进行优化。进一步,编码特征可以作为新的输入,利用同样的方式进行编码,依次循环,最后给网络加上全连接网络和分类器,用少量带标签样本进行微调,形成具有复杂特征提取能力的深度卷积自编码网络。将该网络用于滚动轴承故障识别,将时域振动信号直接输入网络,在公共数据集――西储大学轴承数据集以及实验室实测数据集上均取得了比传统卷积神经网络要好得多的识别效果,例如在实验室实测数据集上将识别精度从0.799提高到了0.921。将底层提取到的特征通过反卷积核逐层重构,第一次在原信号空间看到了神经网络到底“学”到了什么。观察重构信号可知,卷积神经网络对信号特征的提取实际上就是对信号的一种分解,网络底层通道数对应信号分解时基的个数,通道内单个特征对应基分解时的时间点。提出的卷积自编码器以及对网络结构的分析可为后续科研技术人员构建卷积神经网络提供指导。
To solve the problem that the weights of traditional convolution neural network (CNN) can only be initialized randomly, a convolutional auto？？encoder is proposed. The features after convolution and pooling can be reconstructed to the original signal space in the form of stacked deconvolution kernels, in which the stacking weights are the extracted features and the step is the length of pooling. The convolution and deconvolution kernels can be optimized with the goal to minimize the difference between the original and reconstructed signals. Further, after repeating encoding and decoding several times, and adding a fully connected neural network and a classifier to the bottom of the network, a deep convolutional auto？？encoding network is formed. This kind of network with a complex feature extraction ability only needs a few labeled samples for fine？？tune. Applying this network to bearing fault diagnosis, the original time？？domain vibration signal can be input directly into the network without any pre？？processing. On the Case Western Reserve University bearing data set and the measurement data set from laboratory, the proposed method achieves much higher fault recognition accuracy than traditional CNN. For example, the recognition accuracy is improved from 0.799 to 0.921 on laboratory data sets. Reconstructing the extracted features layer by layer through the deconvolution kernel, we can see for the first time what the neural network “has learned” in the original signal space. The process of CNN’s feature extraction is essentially a decomposition of the signal withing the number of the channels is the number of the bases and the feature in one channel corresponds to the time of a base. This conclusion may provide a reference for subsequent research on the construction of convolutional neural network