Out-of-distribution detection-assisted trustworthy machinery fault diagnosis approach with uncertainty-aware deep ensembles
Out-of-distribution detection-assisted trustworthy machinery fault diagnosis approach with uncertainty-aware deep ensembles
复制标题
具有不确定性感知深度集成的分布外检测辅助可信机械故障诊断方法
DOI：
10.1016/j.ress.2022.108648
复制
发表时间：
2022-10-01
影响因子：
8.1
通讯作者：
Li, Yan-Fu
中科院分区：
文献类型：
--
作者：
Han, Te;Li, Yan-Fu
最新的智能故障诊断技术可以有效地识别机械健康状况，同时它们是基于封闭世界假设进行学习的，即训练和测试数据遵循独立同分布（IID）。然而，在现实诊断中，监控的样本往往来自未知的分布，例如看不见的机器故障，从而导致分布外（OOD）问题。这是一个具有挑战性的问题，可能会导致模型对不可预见的机器数据产生不可靠和不安全的决策。为了解决这个问题，开发了一种新颖的OOD检测辅助可信机械故障诊断方法，以提高智能模型的可靠性和安全性。首先，集成多个深度神经网络建立集成诊断系统，称为深度集成。然后，使用不确定性感知深度集成进行可信分析，以检测 OOD 样本并对潜在的不可信诊断发出警告。给出了不确定性阈值的选择准则。最后，通过综合考虑深层集成来实现可信决策？？？预测和不确定性。所提出的值得信赖的故障诊断方法在两个案例研究中得到了验证，在诊断 OOD 样本方面表现出显着的优势。
Recent intelligent fault diagnosis technologies can effectively identify the machinery health condition, while they are learnt based on a closed-world assumption, i.e., the training and testing data follow independently identically distribution (IID). However, in real-world diagnosis, the monitored samples are often from unknown distributions, such as unseen machine faults, leading to an out-of-distribution (OOD) problem. This is a challenging issue that may induce the model to produce unreliable and unsafe decision for unforeseen machine data. To tackle this problem, a novel OOD detection-assisted trustworthy machinery fault diagnosis approach is developed to enhance the reliability and safety of intelligent models. First, multiple deep neural networks are integrated to establish an ensemble diagnosis system, called deep ensembles. Then, the trustworthy analysis with uncertainty-aware deep ensembles is conducted to detect the OOD samples and issue the warnings for the potential untrustworthy diagnosis. A selection criterion of uncertainty threshold is given. Finally, the trustworthy decisions are achieved by comprehensively considering the deep ensembles??? prediction and uncertainty. The proposed trustworthy fault diagnosis approach is validated in two case studies, exhibiting significant advantages for diagnosing OOD samples.