基于深度学习的航空发动机磨损部位识别方法

Aero Engine Wear Location Identification Method Based on Deep Learning

针对航空发动机润滑系统中摩擦副部件复杂、磨损颗粒能谱监测元素众多,靠人工经验难于进行磨损部位精确识别的问题,提出一种基于深度学习的航空发动机润滑系统磨损部位识别方法。该方法应用一维卷积核为计算单元,搭建一维卷积残差网络模型。以航空发动机润滑油中磨损颗粒能谱分析数据为输入,采用所搭建的一维卷积残差网络模型实现对能谱数据的特征提取以及航空发动机磨损部位的定位识别;以某型航空发动机润滑油中磨损颗粒实测能谱数据验证该方法的有效性,并和Resnet18、Resnet34、CNN等网络模型进行对比验证。结果表明,所提方法对航空发动机磨损部位的识别精度达到95%以上。为了验证模型的鲁棒性和泛化能力,在真实的某型航空发动机能谱数据基础上,对含氧数据和噪声数据分别进行测试,进一步说明该模型用于对磨损定位识别的有效性,具备实际应用的可行性。

Because of the complex friction pair components and many wear particle energy spectrum monitoring elements in aero-engine lubrication system,it is difficult to accurately identify the wear parts by manual experience.A wear part identification method of aero-engine lubrication system based on deep learning was proposed.In which a one-dimensional convolution residual network model was established by using one-dimensional convolution kernel as the calculation unit.Taking the energy spectrum analysis data of the wear particles in the aero-engine oil as the input,using the one-dimensional convolution residual network model,the feature extraction of energy spectrum data and the location and recognition of aeroengine wear parts were realized.The effectiveness of the method was verified by the measured energy spectrum data of the wear particles in an aeroengine oil,and compared with resnet18,resnet34,CNN and other network models.The results show that the recognition accuracy of aeroengine wear parts by the proposed method is more than 95%.In order to verify the robustness and generalization ability of the model,the oxygen data and noise data were tested respectively on the basis of the real aero-engine energy spectrum data,which further verified the effectiveness of the model in wear location and identification and the feasibility of practical application.