深度领域自适应及其在跨工况故障诊断中的应用

Deep domain adaptation and its application in fault diagnosis across working conditions

实际生产中,机械设备的工况变化会造成监测数据的分布差异,破坏分类模型的应用基础,降低诊断准确率。为此,提出一种基于深度学习的领域自适应方法,用于跨工况情境下轴承故障诊断。该方法构建两个级联的深度网络:前者用于处理振动信号,自动挖掘故障敏感特征;后者用于将不同工况的样本特征同步映射到一个深度隐藏层(公共特征空间)中,消除工况波动引起的分布差异,生成工况不变特征,实现领域自适应。此外,该深度映射网络可通过参数优化方法自适应构建,能够实现最佳的跨域诊断性能。实验表明,与其他方法和相关研究相比,深度领域自适应在跨工况故障识别中具有更高的准确率。

In real-world production,variable working conditions lead to the distribution discrepancy in monitoring data,which undermines the basis of classification models and consequently deteriorates diagnosis accuracy.Therefore,a deep learning-based domain adaptation method was proposed for rolling bearing fault diagnosis across working conditions.In the method,two deep networks were constructed in a cascaded way.The first one was applied to process vibration signals and automatically mine fault-sensitive features.The second one maps sample features from different working conditions onto a deep hidden layer(the common feature space)to reduce the distribution discrepancy caused by fluctuant working conditions,generate working-condition-independent features,and realize domain adaptation.Moreover,the deep projection network was adaptively established by parameter optimization for the best cross-domain diagnosis performance.Experiments show that,compared with peer approaches and related literature,deep domain adaptation has a higher accuracy in fault recognition across working conditions.