基于浅层Inception-MobileNet旋转机械故障诊断

Fault Diagnosis of Rotating Machinery Based on Shallow Inception-MobileNet

针对现有的旋转机械故障诊断算法存在着时频表示模糊、特征提取困难,从而导致故障诊断效率和精度较低的问题,提出一种基于浅层Inception-MobileNet的旋转机械故障诊断模型。该模型通过拼接法将原始振动信号转换为二维图像,然后采用多尺度卷积核提取不同分辨率的特征图,并结合深度可分离卷积实现特征学习与分类。该网络在CWRU数据集和MFPT数据集上分别实现了十种故障分类和三种故障分类,分类精度为99.5%和95.78%。与传统的网络进行比较,该网络可提高特征提取能力,并且在相同数据集上该网络实现的故障识别精度最高。

The existing fault diagnosis algorithms for rotating machinery have some problems, such as fuzzy timefrequency representation and difficulty in feature extraction, which lead to low fault diagnosis efficiency and accuracy. To solve these problems, a shallow Inception-MobileNet fault diagnosis model for rotating machinery is presented. In this model, the original vibration signals are converted to two-dimensional images by splice method. Then, the multi-scale convolution kernel is used to extract the feature images with different resolutions. The deep separable convolution is used to realize feature learning and classification. The network realizes ten kinds of fault classifications for the CWRU dataset with the accuracy reaching 99.5 % and three kinds of fault classifications for MFPT dataset with the accuracy reaching 95.78 %,respectively. Compared with the traditional network, the proposed network improves the ability of feature extraction and achieves the highest fault identification accuracy for the same data set.