Abstract:

In order to obtain better convergence speed and classification accuracy of bearing fault classification, a Stacked Auto-Encoder（SAE） with dynamic adjustment of learning rate is proposed. In the following iteration, the gradient of the current reconstruction error is used to dynamically adjust the learning rate. According to the positive and negative value of the reconstruction error gradient, two different learning rate reduction strategies are given to make the learning rate more consistent with the current operation of the model. Finally, the accuracy of fault classification and recognition is verified by reverse fine-tuning of different labeled data quantities. The experimental results show that：compared with the fixed learning rate, the dynamic adjusted learning rate reduces 17.70% of the pre training convergence time, 22.92% of the reconstruction error, improves the accuracy of fault classification, and reduces the number of labeled samples on the premise of maintaining the accuracy of classification.

Key words: auto-encoder, deep learning, fault diagnosis, rolling bearing, dynamic learning rate

摘要：

为提高轴承故障分类收敛速度和分类精度，提出一种动态调节学习率的堆叠自编码网络（SAE）。初始时刻给予一个较大的学习率，迭代过程中利用当前重构误差动态调节学习率的大小，根据重构误差梯度的正负值给出两种不同的学习率减小策略，使学习率大小更符合网络当前的运行状态，最后通过不同的有标签数据量进行反向微调，验证故障分类识别的准确率。实验结果表明：相比固定学习率，该动态调节学习率SAE网络预训练收敛时间减少17.70%，重构误差下降22.92%，故障分类准确率得到提高，且能在保持分类准确率的前提下，减少有标签样本量。

关键词: 自编码, 深度学习, 故障诊断, 滚动轴承, 动态学习率