基于双识别器对抗的开放域自适应故障诊断方法

领域自适应问题在机械设备故障诊断领域已被广泛研究,当前大多数封闭域自适应方法通常都假设源域和目标域共享相同的标签类型空间。然而,这不完全符合机械设备真实的诊断需求,实际上会出现新的故障类型,因而传统依据边缘分布对齐的方法难以处理此开放域问题,不能正确辨识出已存在的样本类型和新出现的类型。针对源域和目标域标签类型空间部分重叠的这另一典型开放域诊断问题,提出了一种基于双识别器对抗的开放域自适应故障诊断方法。两个结构相同的神经网络被引入进行对抗性训练,以增强模型对已知类型辨识的领域自适应性能,利用源域与目标域熵最大与最小化策略,以及目标域样本输出的二元交叉方案用以建立分界线隔离新出现的未知类型。利用轴承数据和自吸式离心泵数据进行分析验证,实验结果表明：所提方法相对于典型的封闭域和开放域模型,能更准确的判定机械设备已存在的故障类型和新出现的未知故障类型,在各诊断任务中,均能达到90%以上的正确率。

Open set domain adaptation method based on adversarial dual classifiers for fault diagnosis

The domain adaptation problem has been widely studiedin the field of mechanical equipment fault diagnosis. At present, the most closed set domain adaptation methods generally assume the source domain and target domain share the same label space,which is not practical in real application. This can be called open set domain, because novel fault classes may actually emerge, these conventional methods which only rely on marginal distribution alignment are difficult to separate the new emerging classes and known classes. One of the typical open set domain adaptation problems isthat the label spaces of source domain and target domain are partially overlapped. In this article, a novel open set domain adaptation method based on adversarial dual classifiers(OSDA-ADC)is proposed to address this issue. Two neural networks with the same structure are introduced for adversarial training to enhance the domain adaptive performance of the model for known classes identification. The maximization and minimization entropy strategies of source domain and target domain, as well as the binary cross scheme of target domairl sample output are used to establish a boundary to isolate unknown classes. In addition, the bearing data set and the self-priming centrifugal pump are selected to evaluate the effectiveness of the proposed method. The experimental results show that the proposed method can more accurately identify the existing known fault classes and new emerging unknown fault classes of mechanical equipment than the typical closed set domain and open set domain models. In each diagnosis task,the proposed method can achieve more than 90% accuracy.