基于MKMCC-DSVDD的航空发动机异常检测方法

为解决传统航空发动机异常检测方法准确率和泛化性能较低的问题,提出一种基于混合核最大相关熵的深度支持向量数据描述方法。首先,采用合成少数类过采样技术扩充异常样本规模,提高对非均衡样本的泛化性能;其次,建立基于混合核改进的最大相关熵损失函数,可以在无需数据分布假设的前提下提升准确率;最后,构建基于MKMCC-DSVDD的航空发动机异常检测方法。在某型航空发动机气路系统和滑油系统异常检测实验中,本文方法平均AUC达到98.53%,表明其具有较高的实用性和泛化性能。

Aero engine anomaly detection method based on MKMCC-DSVDD

In order to solve the problems that traditional aero-engine anomaly detection methods are difficult to deal with unbalanced samples, low accuracy and generalization performance, and insufficient data distribution consideration, a deep support vector data description based on mixed kernel maximum correntropy criterion is proposed. First, the synthetic minority oversampling technique (SMOTE) is used to expand the abnormal sample size and improve the generalization performance of non-equilibrium sample. Then, the maximum correlation entropy loss function based on hybrid kernel improvement is established and analyzed to improve the accuracy without data distribution assumption. Finally, an aero engine anomaly detection method based on MKMCC-DSVDD is constructed. The abnormal state average recognition rate reaches 98.53% in the test of anomaly detection of a certain aero-engine gas path system and oil system, which indicates that MKMCC-DSVDD anomaly detection method has high applicability and generalization performance.