基于改进卷积神经网络的变工况轴承故障诊断

Fault Diagnosis for Bearings Under Variable Operating Conditions Based on Improved Convolutional Neural Network

原始信号中的故障特征随工况变化而散布在不同的观测尺度上,针对传统卷积神经网络(CNN)模型仅从单一尺度提取特征,容易出现域移现象并丢失其他尺度信息的问题,提出了基于多尺度自适应加权卷积神经网络(MSAWCNN)的故障诊断模型。首先,采用多个尺度的卷积核并行提取不同观测尺度上的特征;然后,引入自适应加权结构,动态调制多尺度特征以削弱运行条件对特征表达的影响;最后,使用全局均值池化(GAP)层代替全连接层,减少运算量并避免过拟合。利用西安交通大学转速连续变化的轴承数据集进行试验验证的结果表明:MSAWCNN模型的平均准确率达99.69%,具有较强的抗噪性,能从多个尺度全面地提取故障特征,适用于变工况下的轴承故障诊断。

The fault features in original signal are scattered at different observation scales with change of operating conditions.The traditional convolutional neural network(CNN)model only extracts features from a single scale,which is prone to domain shift and loss of other scale information.A fault diagnosis model is proposed based on multi-scale adaptive weighted convolutional neural network(MSAWCNN).Firstly,the multi-scale convolution kernels are used to extract features from multiple observation scales in parallel;Then,an adaptive weighted structure is introduced to dynamically modulate multi-scale features,reducing the impact of operating conditions on feature expression;Finally,the fully connected layer is replaced by global average pooling(GAP)layer to reduce the computational complexity and avoid the overfitting.The experiment is conducted using bearing dataset with continuous changes in rotational speed from Xi’an Jiaotong University,and the results show that the average accuracy of MSAWCNN model reaches 99.69%,having strong noise resistance.The model comprehensively extracts features from multiple scales,which is suitable for bearing fault diagnosis under variable operating conditions.