滚动轴承故障检测深度卷积稀疏自动编码器建模研究
Rolling Bearing Fault Detection using Deep Convolution Automatic Sparse Encoder

针对机械设备故障诊断大多采用有监督学习提取故障特征，而有标签数据难以获取的现状，提出一种在稀疏自动编码器中嵌入卷积网络的深度神经网络。利用希尔伯特和傅里叶变换实现机械设备振动时间序列向Hilbert包络谱的转换，通过卷积网络中多组卷积核自动学习谱空间数据的不同特征，保证了特征提取的自动化、全面性和多样性，稀疏自动编码器搜索具有正交性数据特征的低维表示，并使得编码后的数据具有很强的聚类特性，实现设备的自动故障诊断。通过对滚动轴承振动信号进行分析实验，证明该方法在设备故障诊断中具有去标签化、自动化、鲁棒性等特点。
The supervised learning is commonly adopted in fault feature extraction for mechanical equipment fault diagnosis, while the labeled data are often hard to obtain. To deal with such problem, a deep neural network embedding convolution networks in sparse encoder is proposed. The transformations of Hilbert and Fourier make it possible to transform the vibration time series of machinery into Hilbert envelope spectrum. Different features of spectral space data are automatically learned with multiple sets of convolution kernels in convolution networks, which ensure the automation, comprehensiveness and diversity of the extracted features. The sparse encoder looks for a low-dimensional representation of data featured with orthogonality, making the encoded data characterized by strong clustering; so that the automatic fault diagnosis of the equipment is realized. Through analysis and experiments on vibration signals of rolling bearings, it is proved that this method has the characteristics of de-labeling, automation and robustness in equipment fault diagnosis