Abstract:  Sparse Auto Encoder (SAE) finds a set of "super-complete" base vectors to mine the intrinsic structure and pattern of input data, which enables high-level output to better express the category information of input samples. Its good performance of dimension reduction has been widely concerned and gradually applied in fault diagnosis of mechanical equipment. However, the feature number of hidden layer in SAE model directly affects the expression effect of high-level output on low-level input mode. Simply setting the feature number of hidden layer is difficult to achieve ideal recognition effect. Aiming at this problem, the optimal feature number of each hidden layer is determined by using the advantages of the firefly learning algorithm, and the optimal SAE model is determined. Bearing simulation and fault state recognition experiments show that sparse automatic coding model can achieve better recognition effect than shallow structure and random parameter SAE model under different test samples after the number of hidden layer features is determined, and the recognition accuracy is higher.

Key words:  Deep learning, sparse auto encoder; bearing fault diagnosis, firefly learning

摘要： 稀疏自动编码(Sparse Auto Encoder, SAE)通过寻找一组“超完备”基向量用于挖掘输入数据的内在结构与模式，使得高层输出能够更好的表达输入样本的类别信息，其良好的降维性能受到广泛关注并逐渐应用在机械设备故障诊断中。然而，SAE模型中隐含层特征数直接影响高层输出对低层输入模式的表达效果，简单的设置隐含层特征数难以取得理想的识别效果，针对该问题，利用萤火虫寻优算法的优点，确定各个隐含层的最优特征数，从而确定最优的SAE模型。轴承仿真及故障状态识别实验证明，隐含层特征数确定之后的稀疏自动编码模型在不同测试样本数目下均能取得比浅层结构及随机参数SAE模型更好的识别效果，得到更高的识别正确率。

关键词: 深度学习, 稀疏自动编码, 轴承故障诊断, 萤火虫