航空发动机

蔡舒妤1 ，殷航1 ，史涛1 ，范杰2.基于ResNet-LSTM的航空发动机性能异常检测方法[J].航空发动机,2024,50(1):135-142

基于ResNet-LSTM的航空发动机性能异常检测方法

Aero-Engine Performance Anomaly Detection Method Based on ResNet-LSTM

DOI：

英文关键词:anomaly detection residual neural network long short term memory aero-engine

基金项目:

作者	单位
蔡舒妤1 ，殷航1 ，史涛1 ，范杰2	1.中国民航大学航空工程学院，天津 300300； 2.中国南方航空股份有限公司河南分公司，郑州 450000

摘要点击次数: 302

全文下载次数: 204

中文摘要:

为了实现数据驱动的航空发动机性能异常的智能检测，提出了一种基于残差网络（ResNet）-长短期记忆网络（LSTM）的发动机性能异常检测方法。采用发动机性能数据图像化方法，在数据降维的同时，完备保留数据的关联特征和时序特征；以残差单元构建发动机性能异常检测模型，在加深网络结构的同时，消除深层网络梯度消失问题，提高发动机性能图像空间关联特征的提取能力。同时，引入LSTM，提出基于ResNet-LSTM的发动机性能异常检测模型，通过ResNet与LSTM的融合，强化异常检测模型对时序特征的提取，提升发动机性能异常检测的准确率；通过发动机运行数据进行验证。结果表明：在训练集上，该方法的异常检测准确率为94.95%，比基于ResNet18、ResNet34、ResNet50异常检测模型的分别提高10.87%、8.00%、3.23%；在测试集上，该方法的异常检测准确率为92.15%，比基于ResNet18、ResNet34、ResNet50异常检测模型的分别提高11.81%、9.45%、3.78%。

英文摘要:

In order to realize the intelligent detection of data-driven aero-engine performance anomalies， a method of aero-engine performance anomaly detection based on the Residual Neural Network （ResNet） and Long Short Term Memory （LSTM） is proposed. First， the visualization method of aero-engine performance data is designed. While reducing the data dimension， the correlation features and tim? ing features of data are completely retained. Then， the residual unit is used to construct the aero-engine performance anomaly detection model， while deepening the network structure， the problem of deep network gradient disappearance is eliminated， and the spatial correla? tion feature extraction ability of engine performance images is enhanced. In the meantime， LSTM will be introduced to put forward the model of aero-engine performance anomaly detection based on ResNet-LSTM. Through the integration between ResNet and LSTM， it helps intensify the ability of the anomaly detection model to extract the timing features and enhance the accuracy of this method. Finally， it is verified by the aero-engine operation data. The results show that on the training set， the anomaly detection accuracy of this method is 94.95%， which is 10.87%，8%，and 3.23% higher than that of the anomaly detection model based on ResNet18， ResNet34 and ResNet50， respectively. On the test set， the anomaly detection accuracy of this method is 92.15%， which is 11.81%，9.45%，and 3.78%higher than that of the anomaly detection model based on ResNet18， ResNet34 and ResNet50， respectively.

查看全文查看/发表评论下载PDF阅读器