基于DAL改进CNN的齿轮箱故障诊断及性能分析

Fault Diagnosis and Performance Analysis of Gear Box Based on Dal Improved Cnn Method

为了弥补卷积神经网络(convolutional neural network, CNN)信号容易出现过拟合问题,设计了一种深度主动学习(deep active learning, DAL)改进CNN方法,并成功应用于齿轮箱故障诊断领域。深度主动学习能够将两种方法的优势结合起来,使模型拥有更强的自主学习能力。研究结果表明:完成15次训练后,边缘抽样、深度学习、深度主动学习、随机抽样分别达到75.84%、99.5%、99.79%和93.25%准确率,表明该方法可以有效地过滤出较难鉴别样品。与小波变换相比,傅里叶变换检索精度达到更高97.1%,具有较好稳定性。最大池化正确率达到97.8%,高于均值池化的95.6%。相比较Adam处理器,SGDM处理器在训练集中表现的处理性能略微好一些。该研究可拓宽到其它的机械传动领域,具有很高的应用价值。

In order to compensate for the overfitting problem of convolutional neural network(CNN)signals,a deep active learning(DAL)method to improve CNN is designed and successfully applied in gear box fault diagnosis.Deep active learning can combine the advantages of the two methods to improve the autonomous learning ability.The results show that after 15 training sessions,the accuracy of edge sampling,deep learning,deep active learning and random sampling reaches 75.84%,99.5%,99.79%and 93.25%respectively,indicating that the method can effectively filter out difficult samples.The confusion matrix analysis shows that the DAL algorithm achieves better results than the traditional learning algorithm.Compared with wavelet transform,the retrieval accuracy of Fourier transform is 97.1%higher and has better stability.The correct rate of maximum pooling is 97.8%,higher than 95.6%of average pooling.Compared with Adam processor,the SGDM processor performs slightly better in the training set.The research can be extended to other fields of mechanical transmission and has high application value.