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,...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。展开更多
The nodes number of the hidden layer in a deep learning network is quite difficult to determine with traditional methods. To solve this problem, an improved Kullback-Leibler divergence sparse autoencoder （KL-SAE） is...The nodes number of the hidden layer in a deep learning network is quite difficult to determine with traditional methods. To solve this problem, an improved Kullback-Leibler divergence sparse autoencoder （KL-SAE） is proposed in this paper, which can be applied to battle damage assessment （BDA）. This method can select automatically the hidden layer feature which contributes most to data reconstruction, and abandon the hidden layer feature which contributes least. Therefore, the structure of the network can be modified. In addition, the method can select automatically hidden layer feature without loss of the network prediction accuracy and increase the computation speed. Experiments on University ofCalifomia-Irvine （UCI） data sets and BDA for battle damage data demonstrate that the method outperforms other reference data-driven methods. The following results can be found from this paper. First, the improved KL-SAE regression network can guarantee the prediction accuracy and increase the speed of training networks and prediction. Second, the proposed network can select automatically hidden layer effective feature and modify the structure of the network by optimizing the nodes number of the hidden layer.展开更多
文摘稀疏自动编码(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。
基金Project supported by the National Basic Research Program （973） of China （No. 61331903） and the National Natural Science Foundation of China （Nos. 61175008 and 61673265）
文摘The nodes number of the hidden layer in a deep learning network is quite difficult to determine with traditional methods. To solve this problem, an improved Kullback-Leibler divergence sparse autoencoder （KL-SAE） is proposed in this paper, which can be applied to battle damage assessment （BDA）. This method can select automatically the hidden layer feature which contributes most to data reconstruction, and abandon the hidden layer feature which contributes least. Therefore, the structure of the network can be modified. In addition, the method can select automatically hidden layer feature without loss of the network prediction accuracy and increase the computation speed. Experiments on University ofCalifomia-Irvine （UCI） data sets and BDA for battle damage data demonstrate that the method outperforms other reference data-driven methods. The following results can be found from this paper. First, the improved KL-SAE regression network can guarantee the prediction accuracy and increase the speed of training networks and prediction. Second, the proposed network can select automatically hidden layer effective feature and modify the structure of the network by optimizing the nodes number of the hidden layer.
