基于时间序列异常检测的航空发动机故障诊断  

Aero-engine fault diagnosis based on time series anomaly detection

作　　者：王茵茹 WANG Yinru(AECC Aero Engine Control System Institute,Wuxi 214063,China)

机构地区：[1]中国航发控制系统研究所,无锡214063

出　　处：《航空工程进展》2025年第2期121-133,共13页Advances in Aeronautical Science and Engineering

摘　　要：航空发动机的故障诊断存在数据偏斜问题,即故障样本数量远少于正常样本数量,且故障样本无法反映整个运行工况,导致常规的分类模型泛化能力较差。针对上述问题,提出一种基于改进的深度支持向量数据描述的时间序列异常检测模型。使用长短期记忆(LSTM)网络映射样本的输入和输出,与实际采集输出构成时序异常向量,再通过融入变分自编码器(VAE)的深度支持向量数据描述(DeepSVDD)实现航空发动机时序数据的异常检测;在某型航空发动机地面试车台进行实验验证,与孤立森林(IF)、TranAD(Transformer-based Anomaly Detection)模型及GANomaly等对比方法进行对比。结果表明:采用本文所提模型计算得到特征曲线下面积值达到0.9878,具有最好的异常检测性能,能够有效地应用于航空发动机的各项异常检测及故障诊断任务中。The fault diagnosis of aero-engines is confronted with a data skew issue,where the number of fault sam-ples is significantly fewer than normal samples,and the fault samples can't adequately represent the entire operating conditions,resulting in poor generalization ability of conventional classification models.To overcome this issue,an improved deep support vector data description-based time series anomaly detection model is proposed.The long short-term memory(LSTM)network is employed to map the inputs and outputs of samples,forming temporal anomaly vectors with actual collected outputs.The deep support vector data description(SVDD)incorporating variational auto-encoder(VAE)is utilized to achieve anomaly detection for aero-engine time series data.The ex-perimental verification is performed with a certain type of aero-engine ground test platform,and the model is com-pared to with isolation forest(IF),transformer-based anomaly detection(TranAD)model,and GANomaly.The results show that the curve value calculated with the proposed model can reach to 0.9878,has superior anomaly de-tection performance.The proposed model can effectively be applied to various anomaly detection and fault diagnosis tasks in aero-engine systems.

关 键 词：异常检测 故障诊断 支持向量数据描述 时间序列 航空发动机 

分 类 号：V263.6[航空宇航科学与技术—航空宇航制造工程]