基于GoogLeNet的多状态滚动轴承故障诊断可行性研究

Feasibility Study on Multi-Condition Rolling Bearing Fault Diagnosis Based on GoogLeNet

摘要: 轴承作为轨道交通车辆走行部的关键零部件，由于车辆运行环境复杂，载运工具运行工况不断变化，因此对轴承进行实时的状态监测和智能诊断具有重要意义。本文以深度学习为基础，使用快速谱峭度算法处理轴承振动信号数据，引入GoogLeNet网络进行多状态轴承故障诊断，提出一种基于快速谱峭度的信号–图像转换方法进行多状态滚动轴承数据处理，对比分析了三种数据集和混合数据集下的多状态轴承故障诊断损失和准确度拟合曲线，其准确度均高于77%，由此得出该研究方法具有可行性。

Abstract: Bearings are critical components in the running gear of rail-transit vehicles. Because these vehicles operate in complex environments and under constantly changing service conditions, real-time condition monitoring and intelligent fault diagnosis of bearings are of paramount importance. In this paper, we adopt deep-learning techniques, apply the fast spectral kurtosis algorithm to process bearing-vibration data, and introduce GoogLeNet for multi-state bearing fault diagnosis. A signal-to-image conversion method based on fast spectral kurtosis is proposed to handle multi-state rolling-element-bearing data. Comparative analyses of the loss and accuracy curves obtained with three individual datasets and a combined dataset show that all accuracies exceed 77%, confirming the feasibility of the proposed approach.