基于深度时频特征的机车轴承故障诊断

针对现有机车轴承诊断方法存在故障特征提取不理想、诊断精度低等问题，提出了一种基于深度时频特征的机车轴承故障诊断新方法；利用双通道一维和二维卷积神经网络(CNN)分别对输入的一维原始信号和连续小波变换(CWT)提取的二维时频信号进行深度特征提取；为使输入的一维原始信号简单而有效地反映出信号在时域的全局特征，上通道使用一维CNN，为使输入的二维时频域信号能多角度地反映出信号的细微局部变化，下通道使用二维CNN；在融合层中将上下通道特征自动融合成一个新的深度时频特征，并将提取到的深度融合时频特征经归一化指数函数进行故障分类识别；在此基础上，分析了某局机务段实测的7种机车轴承数据，验证了本文方法的实际工程应用价值。研究结果表明:基于深度时频特征的机车轴承故障诊断方法对7种机车轴承故障的平均诊断精度达到了100%，与一维CNN模型、二维CNN模型和支持向量机(SVM)模型相比，平均诊断精度分别提高了0.7%、1.9%和2.2%；本文方法提取的深度时频特征中每类故障分布间隔规则有序，类内间距很小，而单个一维CNN模型和二维CNN模型提取的特征的每类故障分布间隔不规则，类内间距较大，说明基于深度时频特征的机车轴承故障诊断方法提取深度特征的能力优越，是一种解决机车轴承故障诊断问题的有效模型。 

Locomotive bearing fault diagnosis based on deep time-frequency features

To address the problems such as the unsatisfactory fault feature extraction and low diagnostic accuracy of existing locomotive bearing diagnosis methods, a new method for diagnosing locomotive bearing faults was developed based on the deep time-frequency features. Dual-channel one-dimensional and two-dimensional convolutional neural networks (CNNs) were separately adopted to extract the deep features from the input one-dimensional original and two-dimensional time-frequency signals extracted by the continuous wavelet transform (CWT). A one-dimensional CNN was employed for the upper channel such that the input one-dimensional original signals could effectively reflect the global characteristics of the signals in the time domain. A two-dimensional CNN was applied for the lower channel such that the input two-dimensional time-frequency domain signals could reflect the subtle local changes in the signals from multiple angles. The upper- and lower-channel features were automatically fused in the fusion layer into a new deep time-frequency feature. Then, the extracted deep fusion time-frequency features were classified and identified by a normalized exponential function. Finally, seven types of locomotive bearing data measured in a locomotive depot were analyzed to verify the practical engineering application value of this method. Research results indicate that the average diagnosis accuracies of the proposed method for the seven types of locomotive bearing faults are as high as 100%. Compared with the one-dimensional CNN model, two-dimensional CNN model, and support vector machine (SVM) model, the average diagnosis accuracy of the proposed model increases by 0.7%, 1.9%, and 2.2%, respectively. The distribution intervals of each fault type in the deep time-frequency features are regular and orderly, and the intra-class spacing is very small. Conversely, the features extracted by the single one-dimensional and two-dimensional CNN models exhibit irregular distribution intervals for all fault types, and the intra-class spacing is large. This verifies the superiority of the proposed model in extracting deep features. Therefore, it is an effective model to address the issues in the locomotive bearing fault diagnosis. 4 tabs, 17 figs, 30 refs.