基于多分支深度可分离卷积神经网络的滚动轴承故障诊断研究

Fault diagnosis of rolling bearings based on a multi branch depth separable convolutional neural network

针对传统滚动轴承故障诊断方法存在抗噪性差、需要人工特征提取、计算量较大、对运行设备要求高的问题,提出一种基于多分支深度可分离卷积神经网络(MBDS-CNN)的滚动轴承故障诊断方法,利用深度可分离卷积和权重剪枝技术对模型尺寸进行压缩,通过多分支结构保证模型的精度,避免梯度消失现象的发生。使用模型尺寸、诊断精度、预测速度作为评价指标对模型进行评估。试验结果证明,基于多分支深度可分离卷积神经网络的滚动轴承故障诊断,可以在噪声环境下有效识别轴承不同部位故障程度,提高了诊断效率,降低了对运行设备性能的要求。

Aiming at the disadvantages of traditional rolling bearing fault diagnosis methods,such as poor robust,need for artificial feature extraction,large amount of computation,and high requirements for the running equipment,a fault diagnosis method for rolling bearings based on a multi branch depth seperable convolutional neural network(MBDS-CNN)was proposed.Using the depth separable convolution and weight pruning technology to compress the model size,the multi-branch structure ensures the accuracy of the model and avoids the phenomenon of gradient disappearance.The model was evaluated by using a test set,using the model size,diagnostic accuracy,and prediction speed as evaluation indicators.The experimental results show that the fault diagnosis method for rolling bearings based on the MBDS-CNN can effectively identify the fault degree of different parts of the bearing in the noise environment,improve the diagnostic efficiency and reduce the performance requirements of the running equipment.