基于试飞数据的航空发动机状态监测与故障诊断

Flight Data Based Condition Monitoring and Fault Diagnosis of Aero-Engine

关键词:  航空发动机  飞行试验  状态监测  神经网络  故障诊断

Key words:  Aero engine  Flight test  Condition monitoring  Neural network  Fault diagnosis

摘要:

为了实时监控航空发动机工作参数变化情况，快速及时地预测并诊断发动机故障，本文基于实际试飞数据建立了航空发动机ANN-NARX参数预测模型，考虑到建模样本量大、模型结构复杂、训练时间长、输入输出延迟等因素，采用遗传算法对模型的最小数据样本需求和结构进行了改进优化，并利用蒙特卡洛方法确立了参数预测模型的自适应告警门限，同时，基于构建奇偶空间残差模型实现了航空发动机典型故障诊断。结果表明：实际试飞中只需有限架次试飞数据的训练学习，即可得到发动机参数预测模型，高压转子转速、压气机出口压力、低压涡轮出口温度及滑油回油温度相对误差最大值分别为1.0%，1.7%，0.2%和1.2%，综合模型建模误差和参数测量误差后的自适应告警门限有效降低了模型预测结果的不确定性，在已有数据样本集上的典型故障识别率达到95.2%。

Abstract:

During flying test life cycle， aircraft engine conditions change greatly and faults have been encountered frequently. There always exist urgent needs about monitoring parameters trending on-line， predicting possible faulty condition and diagnosing the specific type when faulty condition encountered. The problem of condition monitoring and fault diagnosis based on flight test data has been studied in this paper. ANN-NARX parameters predicting model of aero engines has been built based on actual flight test data. Considering large demands on data samples， the complex and large design space of ANN model， consequently long training time and input-output time delaying， the model architectures and minimum sample demands have been optimized based on evolving algorithms. The self-adapting thresholds of predicting model have been set using Monte-Carlo method. The specific fault diagnosis has been realized by constructing parity space residual model. All models in this paper have been tested through flight data and applied in actual flying test. The monitoring model could be built based on limited flights in actual flying test. The maximum relative error of high-pressure spool speed， pressure in compressor outlet， total temperature in low-pressure turbine outlet and temperature of all returned oil is 1.0%， 1.7%， 0.2% and 1.2%， respectively. The model predicting uncertainty could be greatly reduced using adaptive thresholds by considering both modeling error and measurement uncertainty. The ratio of detecting and diagnosing specific faulty type is 95.2% based on test samples， which have been encountered in actual flight condition.