Ma Yangfeifei, Zhu Xinyun, Lu Jilong, Yang Pan, Sun Jianzhong
College of Aerospace Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China.
College of Civil Aviation, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China.
Sensors (Basel). 2023 Jul 25;23(15):6660. doi: 10.3390/s23156660.
Anomaly detection and failure prediction of gas turbines is of great importance for ensuring reliable operation. This work presents a novel approach for anomaly detection based on a data-driven performance digital twin of gas turbine engines. The developed digital twin consists of two parts: uncertain performance digital twin (UPDT) and fault detection capability. UPDT is a probabilistic digital representation of the expected performance behavior of real-world gas turbine engines operating under various conditions. Fault detection capability is developed based on detecting UPDT outputs that have low probability under the training distribution. A novel anomaly measure based on the first Wasserstein distance is proposed to characterize the entire flight data, and a threshold can be applied to this measure to detect anomaly flights. The proposed UPDT with uncertainty quantification is trained with the sensor data from an individual physical reality and the outcome of the UPDT is intended to deliver the health assessment and fault detection results to support operation and maintenance decision-making. The proposed method is demonstrated on a real-world dataset from a typical type of commercial turbofan engine and the result shows that the F1 score reaches a maximum of 0.99 with a threshold of 0.45. The case study demonstrated that the proposed novel anomaly detection method can effectively identify the abnormal samples, and it is also possible to isolate anomalous behavior in a single performance signal, which is helpful for further fault diagnosis once an anomaly is detected.
燃气轮机的异常检测和故障预测对于确保可靠运行至关重要。这项工作提出了一种基于燃气轮机发动机数据驱动性能数字孪生的异常检测新方法。所开发的数字孪生由两部分组成：不确定性能数字孪生（UPDT）和故障检测能力。UPDT是在各种条件下运行的实际燃气轮机发动机预期性能行为的概率数字表示。故障检测能力是基于检测在训练分布下概率较低的UPDT输出而开发的。提出了一种基于一阶瓦瑟斯坦距离的新型异常度量来表征整个飞行数据，并且可以将阈值应用于该度量以检测异常飞行。所提出的具有不确定性量化的UPDT使用来自单个物理现实的传感器数据进行训练，并且UPDT的结果旨在提供健康评估和故障检测结果，以支持运行和维护决策。该方法在来自典型商用涡轮风扇发动机的真实数据集上进行了验证，结果表明，在阈值为0.45时，F1分数最高可达0.99。案例研究表明，所提出的新型异常检测方法能够有效识别异常样本，并且还能够在单个性能信号中隔离异常行为，这有助于在检测到异常后进行进一步的故障诊断。
