Abstract:

Aiming at the problems that traditional methods of bearings fault diagnosis heavily depend on expert experience and difficulty in fault feature extraction, combined with the merits of deep learning in dealing with high dimensional and nonlinear data, a novel method based on improved deep wavelet auto-encoder is proposed. Firstly, the loss function of Wavelet Auto-Encoder（WAE） is improved and contractive item constraint is introduced. Secondly, multiple WAEs are stacked to form a Deep Wavelet Auto-Encoder（DWAE） and “cross-layer” connection is introduced to alleviate gradient disappearance. Finally, unsupervised pre-training of DWAE is performed using a large amount of unlabeled data and the model parameters are supervised and fine-tuned with a small amount of tagged data. The bearings diagnosis experimental results show that the method can effectively identify the bearings with multiple fault types and multiple fault severities. The feature extraction ability and recognition ability are superior than Artificial Neural Network（ANN）, Deep Belief Network（DBN）, Deep Auto-Encoder（DAE） and so on.

Key words: fault diagnosis, deep learning, improved wavelet auto-encoder, rolling bearing

摘要：

针对传统滚动轴承故障诊断方法过度依赖专家经验和故障特征提取困难的问题，结合深度学习处理高维、非线性数据的优势，提出一种基于改进深层小波自编码器的轴承智能故障诊断方法。该方法改进小波自编码器的损失函数并引入收缩项限制，再将多个小波自编码器进行堆叠构成深层小波自编码器，并引入“跨层”连接缓解梯度消失现象，最后利用大量无标签数据对网络进行无监督预训练并利用少量带标签数据对模型参数有监督微调。轴承诊断实验结果表明，该方法能有效地对轴承进行多种故障类型和多种故障程度的识别，特征提取能力和识别能力优于人工神经网络、深度信念网络、深度自编码器等方法。

关键词: 故障诊断, 深度学习, 改进小波自编码器, 滚动轴承